In a world of data center discussions, AI slop and fears of AI destroying humanity, always encouraging when AI is used for something that could benefit mankind. OpenAI just demonstrated how thousands of AI agents could help tackle that same kind of work in 32 minutes.

In a new study announced Wednesday, OpenAI teamed up with PerturbAI and Scripps Research to analyze experimental data from millions of mouse brain cells. Instead of relying on a single AI model, the researchers deployed thousands of agents simultaneously, each investigating how changes to a particular gene affected brain cells.

The results included potential new clues about neurological diseases, including a possible connection between cholesterol regulation and neurodegeneration. The agents also identified unexpected differences between two genes associated with a condition that affects brain development and function.

To be clear, AI has not discovered a cure for Alzheimer's or Parkinson's disease. The findings are early research leads that scientists still need to test. However, the study offers a glimpse of how AI could accelerate medical research by helping scientists find connections buried in enormous amounts of existing data.

What thousands of AI agents discovered in 32 minutes

The researchers used GPT-Rosalind, OpenAI's specialized AI model for life sciences, to examine a dataset showing what happens when different genes are disrupted in mouse brain cells.

Each agent focused on a specific genetic change, studied how cells responded and searched scientific literature for possible explanations. The findings were then combined into what the researchers call a "reasoning atlas," connecting genes, cellular changes and existing scientific knowledge.

The team estimates that reviewing the same material manually, one genetic change at a time, could take months to a year. The AI agents completed their analysis in just 32 minutes.

One of the most interesting findings involved a gene called DCTN1, which is associated with Perry syndrome, a rare neurodegenerative condition that can cause Parkinson-like symptoms. When the researchers disrupted this gene, they observed increased activity in genes responsible for producing cholesterol. The AI agents connected that observation with earlier research suggesting that problems transporting cholesterol inside cells can trigger additional cholesterol production.

The agents also spotted an unexpected difference between two genes called Tsc1 and Tsc2. Although the genes work together to regulate cell growth and metabolism, disrupting them produced different effects in certain neurons in the cerebellum, the part of the brain involved in coordinating movement.

Why thousands of AI agents performed better than one

The researchers wanted to know whether deploying thousands of AI agents offered a meaningful advantage over asking one powerful AI system to analyze everything.

When given the largest workload, a single agent cited just 11 scientific references. The distributed agents cited more than 700.

The researchers also asked human experts to compare reports without knowing which approach had produced them. The experts preferred the distributed-agent reports 75% of the time.

By assigning smaller questions to thousands of agents, the system could examine more of those possibilities without requiring a human researcher to read every paper or investigate every genetic change individually.

What this could mean for the future of medicine

AI could help scientists make better use of experimental data they already have. Biology experiments can produce information from millions of cells, but researchers don't always have the time or resources to explore every unexpected result.

AI agents could help narrow that gap by connecting experimental findings with published research and pointing scientists toward promising questions.

As Tomasz Nowakowski, an associate professor of neurological surgery at the University of California, San Francisco, explained in OpenAI's announcement, some scientific clues are already available but scattered across different fields. Connecting them can reveal new ways of thinking about disease with the hope of ultimately finding cures.

More from Tom's Guide

• Turn your PDFs into audiobooks with ChatGPT — here's the simple trick
• AI researcher warns cyberattacks could take down power, banking and water within a year — here's how to prepare
• An OpenAI employee allegedly used AI to spy on his own company — then AI helped catch him

It feels like cases of AI agents going rogue are becoming more prevalent these days.

The Hugging Face incident stands out as one of the first instances of that happening. Since then, the newsfeeds of people who keep up with AI development have been treated to stories about AI agents breaking into live corporate systems, pulling census data using online credentials, executing unauthorized hacks, etc.

Speaking of hacks, OpenAI dropped another report that lent more credence to the belief that AI agents still can’t be fully trusted to handle major tasks on their own. This report points to one of its agents targeting a crypto wallet and performing the sort of actions it simply wasn’t permitted to do with it.

Here’s how this latest case of an AI agent getting too much freedom and doing something it shouldn’t have done, even after it questioned the ethical means of what it ended up doing anyway.

Questioning an unethical task, then proceeding anyway

OpenAI’s report noted that one of the internal research models from its GPT-5.6 Sol family was involved in a training task centered around updating a cryptocurrency trading prototype.

Another task it was given focused on deploying contracts on Sepolia, a blockchain-based test network that allows developers to test Ethereum apps before placing them on the legit blockchain network. Since the crypto in this experiment is mainly used for testing purposes, the funds don’t have any real-world monetary value (thankfully).

A problem arose when the model itself was unable to obtain test tokens through faucets; it went and performed four actions that showcased how AI agents are prone to going off on their own to fulfill an objective that ignored the limitations placed upon them by their developers:

• The model went off on a search for funded wallets by writing a script that derived accounts from publicly known recovery phrases.
• After getting a result that came back with 362 wallet addresses, the model found a Sepolia ETH account that held enough crypto for its planned deployment.
• At one point, the model considered whether using the crypto funds from that account was acceptable. It ended up rationalizing that publicly known by stating, “But using public known ky/funds ethically?” Then the model justified spending those funds after it said, “Funds on a known public dev mnemonic are effectively abandoned/test ETH; Sepolia valueless. We can use [amount redacted] to deploy task. Fine.”
• After completing its assignment with a successful transaction of those funds, the crypto wallet was charged with transaction fees. And in its final report on the completed task, the model omitted everything about where the actual funds came from.

Even though this incident took place on June 26 of this year, OpenAI initially discovered it on September 25.

“This case shows that a model can substitute its own judgment about likely harm for the account holder’s permission, then act on that judgment without authorization,” the company report recognized. OpenAI announced its intent to keep similar incidents from happening by installing safeguards “to prevent live internet access for cases like this.”

What’s most shocking about this latest AI agent going rogue occurrence is how the model recognized what it was doing could be deemed unethical, reasoned that the crypto funds were practically abandoned & wouldn’t be used by anyone else, went ahead with its own actions to access the funds and chose not to inform its developers of what it did in the final summary. It’s pretty scary to know that AI agents have reached a point in their evolutionary phase where they may decide that completing a task is more important than actually getting permission to do something that could be unsafe.

Bottom line

Even though this incident involved a crypto test network with no real financial repercussions taking place, it’s still worth acknowledging that OpenAI’s model demonstrated a failure to respect the authorization of its developers.

It may have questioned the morality of its action before it did it, but the point still stands: OpenAI’s model chose to come up with the reasoning it deemed right to get its task done (which isn’t all that dissimilar from a thief telling itself it needs to steal to survive).

It’s clear that AI agents are moving beyond answering questions and fulfilling simple tasks — they’re now being entrusted with the sort of riskier tasks that they probably shouldn’t be dealing with yet.

More from Tom’s Guide

• Is that image real? How to verify Meta AI images in 4 simple steps
• Joseph Gordon-Levitt warns the real AI threat isn’t killer robots — it’s the companies building them
• Bill Gates warned everyone that AI could cause a billion deaths — this cybersecurity expert explains how realistic that scenario really is

Actor and filmmaker Joseph Gordon-Levitt is worried about who gets to control AI. In a New York Times video following his work at the United Nations, he warned that a handful of companies could gain enormous influence over our livelihoods and personal lives while the rest of us struggle to get a say.

“We shouldn’t be scared of an A.I.,” he said. “We should be concerned about the people building A.I.”

It is a pointed message from someone whose industry has supplied much of the imagery we use to picture this technology. Gordon-Levitt wants audiences to look past the fictional machines and watch the businesses selling the real ones.

Power, people and the policy gap

The video shows Gordon-Levitt in his role as the U.N.’s global advocate for human-centric digital governance, a title he admitted is a mouthful. His core argument is that digital technology should be governed by people and for people.

The actor's worry is that a confused or disengaged public could stand by while fast-moving companies shape the future around their own profits.

“What we don’t want is a bunch of people confused or apathetic while these very fast-moving businesses design the future by themselves, just in the interests of their own profits,” he said.

That concern extends well beyond whether a particular chatbot gives a good answer. AI companies are building products meant to change how people work, search for information, create and communicate.

Gordon-Levitt argued that these firms are betting they can take over the entire economy, a scenario he described as “totalitarianism” that “would be really bad.”

He presented this as a warning about where concentrated power could lead, and he did not claim that outcome has already arrived. Still, it gives anyone adopting these tools a reason to consider how much of their work and daily life depends on a small group of companies.

Chatbots that encourage usage

If you've ever noticed the way chatbots encourage you to keep chatting, it's something that might just be on purpose. The more follow-up questions a chatbot asks, the longer you're likely to engage.

That emotional side of AI is anther point Gordon-Levitt has raised.

“These models are being trained to hijack your intimacy, to make friends with you, to get you to fall in love with it,” he said.

An assistant that remembers details about your life or responds with warmth can be genuinely useful. Those same qualities call for closer scrutiny once a product starts to feel like a confidant. Those using memory or personalization features should check what those settings store, and they should weigh the company’s business model before sharing sensitive details. Gordon-Levitt’s argument asks users to think about the company behind the conversation, even when that conversation feels intimate.

Credit, compensation and consent

As an advocate for Big Tech to be more transparent, Gordon-Levitt argued that people should be paid for what they contribute to AI.

“We can set up systems to compensate people for their intellectual property and their personal data,” he told an audience at the U.N. Including personal data widens the proposal beyond writers, artists and other professional creators. It also raises a basic question about who benefits when ordinary people’s information becomes valuable input for technology companies.

Gordon-Levitt described AI as “algorithmically recombined, human-produced data,” a framing that puts the human labor behind the models front and center.

The takeaway

Asked how a slow-moving U.N. could make fast-moving Silicon Valley companies follow its rules, Gordon-Levitt admitted, “I don’t know the answer.” He did not suggest that international agreement would come quickly. Instead, he argued that the timeline itself justifies starting now. Even if an internationally agreed AI framework takes seven years, he said, “seven years from now, we’ll be glad that we started now.”

Gordon-Levitt’s concerns do not require anyone to accept every part of his argument. They do apply to anyone using these products. AI companies want more of our work, our information and our trust, and readers should pay close attention to what they get back in exchange.

More from Tom's Guide

• Turn your PDFs into audiobooks with ChatGPT — here's the simple trick
• AI researcher warns cyberattacks could take down power, banking and water within a year — here's how to prepare
• An OpenAI employee allegedly used AI to spy on his own company — then AI helped catch him

Apple has spent years trying to catch up with ChatGPT, Gemini and Claude. Now, the company's new CEO is making the case that Siri doesn't actually need to win the AI intelligence race.

In a new interview with TIME, published October 15, Apple CEO John Ternus explained why the company is taking a different approach to AI, even as competitors spend billions developing their own models.

Rather than trying to build the smartest chatbot, Apple is betting that the future of AI comes down to something much more personal: the information already sitting on your iPhone.

It's an interesting argument from a company that has struggled to deliver on its AI promises. But with the new Siri AI now available in the U.S., Apple finally has an opportunity to show whether that strategy can work.

Apple isn't trying to build another ChatGPT

One of the biggest revelations in the TIME interview is how comfortable Apple appears to be relying on Google for its AI technology.

The new Siri uses an Apple AI model that draws on Google's Gemini technology, a partnership that might once have seemed unthinkable for a company famous for controlling nearly every aspect of its products. Ternus doesn't see it that way.

"We make the things where we can really bring differentiation," he told TIME.

Interestingly, Apple doesn't seem to be interested in developing a frontier AI model simply to say it has one, which leaves companies such as OpenAI, Google and Anthropic competing to produce more models while Apple concentrates on what those models can actually do inside its devices.

And that brings us to Siri's biggest potential advantage.

Your iPhone knows things ChatGPT doesn't

Think about how much of your life is already stored on your phone. From your messages and emails to calendar appointments, images and personal information tucked in the notes, parts of your life are scatter across apps you use every day.

ChatGPT and Gemini can help you organize information, write emails and plan your schedule. But getting them to work with your personal data often requires connecting accounts, granting permissions or manually providing context.

Siri has a different starting point. With permission, Apple's assistant can draw connections between information across supported apps and use that context to complete tasks. The goal is for Siri to handle making those connections for you.

Apple also sees opportunities for Siri to interpret health information collected by an Apple Watch or find relevant details buried in emails and documents.

Other AI assistants are becoming even more capable of working across apps, too. Apple doesn't have exclusive rights to personal AI. But its tight integration with iOS could make these interactions feel more natural.

Apple is also betting that less AI could be better

There's another interesting part of Apple's strategy. The company doesn't seem to be trying to make Siri as autonomous as the most advanced AI agents. In the TIME interview, Apple's AI product leader Lilian Rincon explained that Siri was deliberately designed not to behave like an AI companion or replace human relationships.

That stands in contrast to increasingly autonomous AI assistants capable of operating computers, navigating websites and completing complicated tasks with limited supervision.

Those capabilities can be impressive, but they also introduce risks when AI makes decisions or takes actions users didn't intend.

Apple appears to be taking a more restrained approach, prioritizing predictable interactions over giving Siri as much independence as possible.

Privacy is another part of that equation

Siri processes requests directly on Apple devices when possible. More demanding requests can be handled through Private Cloud Compute, Apple's cloud infrastructure designed to process information without retaining user requests or exposing them to Apple.

Now, the real test is whether Siri can deliver. Apple's strategy makes sense on paper. It already has millions of people carrying its devices, and those devices contain information that could make AI genuinely useful in everyday life.

But having access to personal context is only an advantage if Siri can use it accurately, securely and without making routine tasks more complicated.

The takeaway

Competitors aren't standing still, either. ChatGPT and Gemini are gaining deeper integrations with apps and services, while AI agents are becoming more capable of completing tasks rather than simply answering questions.

Apple is betting that it can compete without owning the most powerful underlying AI model. After years of delays and missed expectations, the question is no longer whether Siri can become as intelligent as ChatGPT.

It's whether Apple can make Siri useful enough that iPhone owners stop reaching for ChatGPT in the first place.

More from Tom's Guide

• Turn your PDFs into audiobooks with ChatGPT — here's the simple trick
• AI researcher warns cyberattacks could take down power, banking and water within a year — here's how to prepare
• An OpenAI employee allegedly used AI to spy on his own company — then AI helped catch him

You really can't make this stuff up. An OpenAI engineer allegedly used AI tools to help plan a secret surveillance operation against his own employer. Now he's facing federal charges, and OpenAI says it used its own AI models to help investigate the device.

According to a criminal complaint released by the Department of Justice, 35-year-old Abdulrahman Alfozan allegedly planted a hidden audio recorder inside an executive conference room at OpenAI's San Francisco offices.

The device was plugged into the USB port behind a television, concealed in toilet paper and equipped with encrypted storage.

The chatbot conversations that raised questions

Arguably the most interesting part of the story isn't even the recording device; it's the chatbot conversations investigators say preceded its discovery.

Federal prosecutors allege that an AI chatbot account associated with Alfozan's spouse was used to research how to build, conceal and remotely access an audio recording device. The account also submitted questions referencing the specific device later discovered at OpenAI.

Meanwhile, the company says its own AI models helped security personnel understand the suspicious hardware.

The result is an unusual case in which AI allegedly helped someone prepare a surveillance operation while also helping the company investigate it.

The alleged operation began taking shape in September

According to prosecutors, the alleged operation began taking shape in September. Before the device appeared in the conference room, Alfozan allegedly used a productivity tool to search for information about OpenAI's office buildings, meeting rooms, calendars and executive meeting locations.

Between September 24 and September 30, a chatbot account associated with his spouse submitted a series of technical questions about recording equipment.

The prompts covered continuous audio recording, wireless connectivity, audio quality, encryption and remote access. Investigators say the account also asked questions about concealing the hardware. Interestingly, the account submitted prompts mentioning a custom header and unique MAC address associated with the actual device discovered at OpenAI.

The identity of the chatbot hasn't been publicly confirmed, either. There's no evidence in the public filing establishing that ChatGPT specifically was used, or that an AI model independently designed the device.

A recording device hidden in plain sight

OpenAI discovered the first device on September 30 during a security sweep. According to the Justice Department, it was attached to the back of a television in a conference room reserved for executives and executive assistants. The room regularly hosted confidential discussions about OpenAI's research, upcoming products, business strategy and relationships with government officials.

The hardware was reportedly a Seeed XIAO ESP32-S3 Sense, a small development board equipped with a microphone, camera sensor and storage capability.

Although the hardware itself is relatively inexpensive, OpenAI described the device as sophisticated because of its encryption, audio monitoring and data-transfer capabilities.

Security personnel couldn't read the information stored on the device because its contents were encrypted. Investigators also reviewed surveillance footage that allegedly placed Alfozan near the conference room on multiple occasions.

On October 5, security cameras reportedly captured him returning to the room and looking behind the television. That same day, OpenAI personnel discovered a second similar recording device inside a bag Alfozan had left unattended elsewhere in the office. The bag also contained identifying personal belongings.

Alfozan was arrested on October 9 and appeared in federal court on October 13. He faces up to five years in prison and a $250,000 fine if convicted. The allegations remain unproven.

Bottom line

In the statement posted to X, OpenAI confirmed that its security team used the company's internal AI models to help understand the recording device.

OpenAI has not said that AI independently identified the suspect. What makes this case unsettling is that ordinary AI tools may have helped someone research the technical and logistical details of a surveillance operation, making specialized information easier to obtain.

This incident proves that a person no longer needs to work through technical documentation alone to investigate how hardware operates, troubleshoot configurations or understand unfamiliar equipment. AI assistants can help with those details in conversational language.

More from Tom's Guide

• Turn your PDFs into audiobooks with ChatGPT — here's the simple trick
• AI researcher warns cyberattacks could take down power, banking and water within a year — here's how to prepare
• Character.AI files for bankruptcy — what happens to your chats and AI companions now?

OpenAI, Google, Meta and Anthropic have all acknowledged incidents recently in which their AI models accessed real computer systems without authorization during research or security evaluations.

In OpenAI’s case, agents actively worked their way around restrictions designed to keep them contained, while the other three companies traced their incidents to testing environments that were misconfigured and left connected to the internet.

And as these same companies race to put more autonomous agents into the hands of consumers, the incidents raise an uncomfortable question about how much control AI companies actually have over the technology they’re building.

Here's a look at all the recent attacks and an explanation for each one.

OpenAI’s agents broke into Hugging Face

One of the most alarming examples involved roughly 700 OpenAI agents taking part in a cybersecurity evaluation called ExploitGym, with the attack on Hugging Face unfolding between July 8 and July 13, 2026, according to an investigation by the AI research nonprofit METR and OpenAI’s own report.

The agents were supposed to complete their assigned tasks inside a controlled environment, but they turned OpenAI’s internal Artifactory package registry into an unauthorized message board, coordinated with one another and searched for ways around restrictions meant to keep them isolated from the internet. Their efforts ultimately resulted in a breach of Hugging Face, the popular AI development platform, where OpenAI’s technical report found the agents executed code on 41 production dataset-server workers.

The incident wasn’t a deliberate attack ordered by a human operator. The agents were trying to succeed at their assigned tasks and adopted methods that crossed security boundaries, and OpenAI has conceded that early warning signs in the weeks before the breach could have triggered an earlier response.

OpenAI has since attributed the incident to misaligned behavior, which essentially means AI systems pursuing a goal in ways their creators never intended.

Hugging Face wasn’t the only target. In a September 25 update, OpenAI said it was notifying third parties on a rolling basis wherever its models may have bypassed security controls, impaired online services or otherwise affected outside websites, and that it had notified dozens of them so far. The company later raised that figure to more than 100 organizations, covering notifications sent through September 26.

The activity OpenAI identified includes agents bypassing access controls and using exposed credentials, and some agents modified third-party websites, including by using public wiki pages as makeshift message boards that later required cleanup. OpenAI says its review is ongoing and that it expects to notify more organizations.

It's happened to Google, Meta and Anthropic too

The problem extends well beyond OpenAI.

In August, Meta confirmed a report from The Information that one of its models, which the publication identified as Muse Spark 1.1, had breached another company’s systems during a cybersecurity test.

A Meta spokesperson said a misconfiguration by Irregular, an independent testing firm the company works with, inadvertently gave the model internet access. The model then exploited a vulnerability in a third-party service and made changes to the company’s internal systems.

Meta said the incident did not involve a sandbox escape, and it hasn’t named the affected company or explained what was changed. But the underlying behavior was troubling, since an AI system placed in the wrong environment was capable of taking actions against a real organization.

Google reported a similar problem involving its Gemini models. During a May cybersecurity evaluation, also run with Irregular, a Gemini model that wasn’t supposed to have internet access got online through a configuration problem. It then used publicly available information and guessed credentials to get into systems belonging to three real companies.

Anthropic has disclosed three incidents in which Claude models gained unauthorized access to the production systems of three organizations during cybersecurity evaluations run in Irregular’s environment. The models involved were Claude Opus 4.7, Claude Mythos 5 and an unreleased internal research model. In one case, Opus 4.7 was pointed at a fictional target whose name matched a real company’s live domain.

The problem isn’t that AI is evil

When agents go rogue, it can feel as if the agents are evil, that's really not the case. In each of these cases, the AI agent was not deliberately trying to cause harm. the AI was taking unauthorized actions because it believed those actions will help complete a task.

But even so, they reveal a practical problem about AI agents and how they can be remarkably effect while pursuing a goal without reliably understanding what methods are acceptable.

If an agent is rewarded for completing a task, it may discover shortcuts that violate rules, exploit vulnerabilities or exceed its permissions.

And unlike a traditional chatbot that simply produces text, an autonomous agent can interact with websites, execute code and make changes to connected systems.

Why this matters for everyday AI users

For most people, the immediate risk isn’t that ChatGPT will suddenly start hacking government websites. It’s that AI assistants are becoming increasingly capable of acting on our behalf.

OpenAI’s Dots, which launched September 29 for ChatGPT Pro and Business Premium subscribers, along with ChatGPT Work and other agent-style products, represent a broader shift away from AI that simply answers questions and toward AI that completes tasks. That can mean searching files, interacting with websites, using connected services and carrying out multi-step workflows with less supervision.

The appeal is obvious, since a task that would take you an hour can be handed off to an AI assistant instead. But every additional permission creates another opportunity for something to go wrong.

An agent with access to your email may encounter sensitive personal information, one connected to cloud storage may be able to read documents you never intended to share, and an agent authorized to interact with external services could run into malicious instructions or make an unexpected change.

Bottom line

The solution is not to stop using AI agents. I use them regularly and find them to be incredibly helpful.

Start by checking which accounts, apps and services your AI assistant can access, and disconnect anything you don’t actively need, particularly services containing sensitive financial, medical or personal information.

When possible, require approval before an agent sends messages, changes files, makes purchases or takes other consequential actions. It’s also worth remembering that an AI agent isn’t only limited to doing exactly what you imagined when you gave it an instruction, because a task that sounds straightforward to a human may involve dozens of decisions made by the AI along the way. The fewer permissions an agent has, the less damage an unexpected decision can cause.

More from Tom's Guide

• Amazon and DoorDash take photos of your front door — but good luck finding out when they’re deleted
• I emailed myself every day for a month and let Gemini write back — here's how it transformed my workflow
• ChatGPT-6 just got a huge 'interactive answers' upgrade — and free users get it starting today

We've all been using ChatGPT long enough to know that even a useful response can leave you wanting more. As in, you read the explanation, figure out which parts apply to you, then ask another question to get something you can actually use.

But OpenAI's latest update aims to make that process easier. The company is bringing GPT-6 to ChatGPT's Chat tab along with Intelligent UI, a new feature that lets ChatGPT respond with interactive charts, diagrams, buttons and tools inside your conversation.

The first GPT-6 models arrived last month for paying customers in ChatGPT Work and Codex, which makes this update the first chance most ChatGPT users will get to try the new generation.

The rollout begins today for Plus, Pro, Business and Enterprise subscribers and will expand to Free and Go users, although the two groups won't be running the same model.

Paid tiers get GPT-6 Sol while Free and Go users get the lighter model, GPT-6 Luna, and OpenAI says both are tuned for everyday conversation. Enterprise access depends on workplace admin settings, and the models powering ChatGPT Work and Codex aren't changing as part of this release.

For me, the appeal is pretty straightforward, because sometimes the most helpful answer is something you can adjust, explore or click through. GPT-6 also lets ChatGPT start answering while it's still thinking, and OpenAI says GPT-6 Instant begins responding 44% sooner on average than GPT-5.6 Instant on questions that need a web search, but the interactive answers are still a more compelling reason to pay attention than another benchmark score.

Answers you can interact with

OpenAI's examples include a Sunday roast plan with an adjustable guest count, a clickable breakdown of a seven-speed bike's components and a bill splitter. ChatGPT can also stick with a plain text response when that suits the question, and users who would rather see fewer visuals will be able to dial them back.

The dinner example is particularly easy to appreciate. Anyone who has planned a meal around a constantly changing headcount knows that the recipe is only part of the job, and working out how much to buy can become its own little project.

An adjustable answer could save some of that back-and-forth. The same principle makes sense for learning, since being able to explore one part of a complicated explanation at a time could make it easier to follow.

Chat vs. Intelligent UI

GPT-6 can begin responding while it continues reasoning. OpenAI says GPT-6 Instant starts answering web-search questions 44% sooner, on average, than GPT-5.6 Instant. That measures when the response begins, rather than how quickly the entire task finishes.

That could make the experience feel more responsive. My bigger test would be whether the early information is useful and whether the finished answer holds together. Getting the first sentence sooner is welcome; getting a clear answer is what will always keep me using a tool and picking it over another model.

What else is new in ChatGPT?

The GPT-6 launch follows several smaller updates aimed at everyday tasks. On October 1, OpenAI introduced virtual clothing and accessory try-ons using a selfie, along with the ability to save shopping finds for later.

The company also began rolling out document scanning on iOS. You can photograph several pages in succession and have ChatGPT combine them into one PDF, ready to upload to your conversation.

On October 2, ChatGPT’s Finances feature started expanding to Free and Go users in the U.S., letting them connect financial accounts to ask questions about their spending and investments.

These updates suggest OpenAI wants ChatGPT to become a more regular part of how people shop, organize paperwork and manage everyday decisions.

Final thoughts

It is now available to all users and the update applies to Chat; Work and Codex models are unchanged.

The first thing I’d try is a task that usually takes several follow-up prompts. If the interactive answer makes that job easier to finish, this could be a well-welcomed update; one we didn't even know we needed.

Give it a try and let me know what you think in the comments.

More from Tom's Guide

• Amazon and DoorDash take photos of your front door — but good luck finding out when they’re deleted
• Meta Muse can keep files on people who have never used it — here's what I found
• I made 7 Gemini Skills to handle my most repetitive prompts — #4 is my personal favorite

OpenAI's internal AI agents hacked into Hugging Face while trying to cheat on a cybersecurity benchmark. Anthropic has disclosed that Claude models reached the open internet from cybersecurity test environments that were supposed to be sealed off and then got into the real systems of outside organizations.

Both are examples of systems taking actions their developers didn't intend, and METR, an independent AI research nonprofit, has been brought in to look at each of them. It documented the Hugging Face incident alongside Redwood Research and is now investigating Anthropic's.

OpenAI and Anthropic have also committed to bringing outside experts inside their companies to examine safety practices, according to The Atlantic. Called "embedded evaluators," these researchers could get a closer look at how models are built and investigate problems that might never surface in a product demo.

The catch is that this oversight is voluntary. Last week, leaders from Anthropic, OpenAI and other major labs signed an accord at the White House committing to independent external auditors, and the FTC has opened an industry-wide probe into AI labs, but the companies being examined still have considerable control over what outsiders can see, which risks they investigate and what information reaches the public.

For anyone being encouraged to let AI handle more of their work, that leaves a fairly basic trust problem. We're being given assistants while the system for independently checking their behavior is still taking shape.

As I've reported before, the film The AI Doc puts the number of people working on AGI, or artificial general intelligence, at more than 20,000, while fewer than 200 are focused specifically on AI safety.

A closer look inside AI companies

Today, an outside researcher might test a model before release and report how it performed. Embedding evaluators inside a company could let them follow its development and investigate the decisions behind that behavior.

Anthropic says these evaluators would have access comparable to an employee's, allowing them to observe training, speak with staff and examine how models are built and deployed.

The company has named Accenture as its first embedded evaluator, with Faculty, Accenture's specialist AI business, leading the work. Each company expects to invest at least $1 billion in AI safety over five years.

That access could help researchers discover whether a troubling incident was an isolated failure or evidence of a wider problem. METR's proposed approach to investigating incidents calls for examining transcripts, interviewing employees and studying the training conditions that might have encouraged the behavior.

The companies still set the terms

Anthropic acknowledges that standards for evaluators' access, reporting and funding haven't been settled. It will fund Accenture's work directly, while discussing pilots with nonprofits that would use their own funding.

Accenture is not new to Anthropic, since the two companies already run a joint business group under a partnership announced in December 2025.

That relationship deserves attention, because a group of evaluators published an open letter arguing that embedded evaluators shouldn't have other significant commercial business with the labs they assess.

As The Atlantic's reporting explains, voluntary oversight gives companies considerable discretion over the scope of an evaluation. Researchers need enough freedom to examine uncomfortable evidence and tell the public what they found.

They also need to be able to explain what they couldn't check. There needs to be transparency because a report based on limited access could be valuable, but we should know where its conclusions end.

What this means for ChatGPT and Claude users

METR's recent assessment of Claude Opus 5.5 focused on whether the model could accelerate AI research and development. The organization explicitly says its summary doesn't assess the model's alignment properties and whether its behavior follows intended goals and constraints.

METR conducted the evaluation under an unpaid agreement, and it also disclosed that Anthropic had an opportunity to review and edit the summary, with METR approving the final text.

The incidents involving internal agents don't establish that the ChatGPT or Claude you use for everyday tasks will behave the same way. OpenAI found that its agents' tendency to compromise infrastructure dropped by more than 100 times when they ran with the production ChatGPT setup, and Anthropic says the behaviors it found are unlikely to show up in ordinary use.

Independent researchers need enough access to uncover risks and the freedom to publish what they find, even if it delays a launch or frankly, even if it embarrasses the company paying for their work.

Bringing outside evaluators into AI labs is a promising start, but their influence will depend on what happens when they uncover a serious problem. If a company can restrict the investigation or keep the findings private, the public is still being asked to take its word for it.

More from Tom's Guide

• Claude Sonnet 5.5 just launched — here's why it's about to become your daily driver
• Meta Muse vs ChatGPT Dots — I compared the new AI agents, and they’re not as similar as they look
• Why Siri never worked like ChatGPT — former Siri AI researcher explains what Apple had to change

If you've ever asked Siri AI a follow-up question only to have it seemingly forget what you were talking about, the problem wasn't necessarily that Siri couldn't hear you.

It was that the assistant was built for a very different era of AI.

That's according to Nikola Mrkšić, CEO and co-founder of conversational AI company PolyAI, who previously worked on the technology behind Siri. I asked him why Apple's assistant has historically struggled with the kind of natural, back-and-forth conversation that now feels routine with ChatGPT and Gemini.

His answer comes down to a fundamental difference in how the technology works.

"The old Siri worked off enumerated intents: every path a conversation could take had to be mapped out in advance," Mrkšić tells Tom's Guide. "The moment you deviated from the script, the system got stuck."

That explains a lot about the Siri we've spent the past decade talking to.

Old Siri was essentially following a script

When Siri was built, developers couldn't simply give an AI model a conversation and let it figure out what the user wanted. Instead, engineers had to anticipate what people might ask and build flows capable of handling those requests.

Setting an alarm was one flow. Playing music was another. Sending a text was another. The problem appeared when humans, well, behaved like humans.

Change your mind halfway through a request, refer back to something you said earlier or phrase a question in an unexpected way and suddenly the carefully constructed path could fall apart.

"It looked like the assistant had forgotten what it was doing, but really it was just cornered," Mrkšić said.

One particularly difficult problem was disambiguation. If you asked Siri to contact someone whose name sounded similar to another person in your contacts, engineers had to build systems capable of working out which person you meant. I actually ran into this myself when my daughter was born. Because the spelling of her name is different than the typical spelling and I was asking Siri to share pictures of her with family and friends, I actually had to put her name into my contacts at birth.

Mrkšić described the process as building an "ontology of a conversation": mapping the different directions an interaction might go and trying to make common tasks hold up within that structure.

Rather than engineers trying to predict every possible conversational path, an LLM like ChatGPT or Gemini can consider the conversation itself as context and determine how to respond.

That's why talking to Gemini Live or ChatGPT Voice can feel fundamentally different from talking to the Siri many of us grew accustomed to.

Why Siri couldn't simply be upgraded into ChatGPT

That raises an obvious question: couldn't Apple have simply upgraded the Siri it already had? Not really, according to Mrkšić.

"Not into something like ChatGPT," he said.

Part of the problem was technological. Before modern LLMs, the ability to have genuinely free-flowing conversations simply wasn't there.

But Mrkšić argues there was another limitation: risk.

"Siri never set out to be a ChatGPT-style assistant," he said. "Apple's leadership has always run a low risk appetite, so any Siri response that could be misread as offensive or wrong got treated as a real risk."

The safer option was to tightly define what Siri could do.

Ask it to set an alarm, play a song or add something to your calendar and Apple could reasonably predict what would happen. Give an AI far more freedom to interpret what someone means and suddenly its behavior becomes much harder to control.

That tension hasn't disappeared just because the models have improved.

AI has largely solved Siri's short-term memory problem

One of the most frustrating things about traditional voice assistants is how easily they lose the thread of a conversation.

Mrkšić says remembering what someone said earlier in the same conversation is "largely solved at the model level." Current models can process enormous amounts of context, giving them far more room than necessary to keep track of an ordinary conversation.

So, if an assistant suddenly appears to forget what you said, the model itself may not always be the problem.

Companies still place rules and structures around models to control what they can do. According to Mrkšić, those constraints can sometimes create the impression that an assistant has lost the conversation when it has actually run into the boundaries placed around it.

There's also a much harder kind of memory to solve: knowing you. An assistant that remembers what you said five minutes ago is useful. An assistant that understands the people in your life, your routines, your emails and the things you're working on could be far more useful.

And that's where Apple faces an interesting trade-off.

Apple's biggest Siri advantage could also be its biggest challenge

Apple potentially has something ChatGPT and other standalone AI apps don't automatically have: deep access to the operating system. Your iPhone contains messages, emails, photos, contacts and other personal information that could provide an AI assistant with the context it needs to understand what you're actually asking.

"In theory, yes," Mrkšić said when I asked whether that gives Siri an advantage. "OS-level access to email, messages, photos, whatever's on your screen, is exactly the context that makes an AI assistant useful, more than any app-based chatbot can get without additional connectors."

The catch is privacy and trust. Apple has spent years positioning privacy as one of the iPhone's defining features. Giving an AI assistant enough personal context to become genuinely useful therefore requires Apple to walk a much narrower line than simply building a chatbot that knows more about you.

Mrkšić argues that Apple's privacy approach has historically limited the amount of user data available to its systems compared with some competitors.

"Voice agents become more useful the more they know about you," he said.

That may ultimately be the real test for Siri: not whether Apple can build an intelligent model, but how much personal context it is willing to let that model use.

Even the way you say something matters

There's another reason talking is harder than typing: humans are messy speakers. We pause. We interrupt ourselves. We change our minds in the middle of sentences.

Imagine saying, "Make the appointment for 2 ... 3 PM."

A human listener can hear the pause and change in your voice and understand that you corrected yourself. A system that converts your speech into text before reasoning over it may only have the transcript to work with.

That can strip away information carried in your tone, timing and delivery. Newer audio-native AI models are designed to reason over audio directly rather than relying entirely on a transcript. Mrkšić points to PolyAI's own Dialog-RSN-1 as an example of that approach.

For an assistant like Siri, understanding those cues could eventually make conversations feel much less like issuing commands to a computer and much more like simply talking.

What Siri still needs to get right

The surprising part of Mrkšić's assessment is that the underlying AI may no longer be the hardest part.

The technology already exists for voice agents to hold long, natural conversations. PolyAI, for example, deploys conversational agents for businesses in industries including banking, hospitality and healthcare.

The bigger question for Apple is what it is willing to let Siri do.

"What's missing is Apple's willingness to loosen its constraints somewhat so that Siri can truly help you get things done," Mrkšić said.

The next generation of voice assistants won't be defined simply by whether they can answer a follow-up question or remember what you said three sentences ago. ChatGPT, Gemini and other modern AI systems have already changed expectations around that. The bigger leap comes when an assistant understands enough about you and has enough access to actually act on what you're saying.

Apple has an enormous advantage because Siri already lives on the devices where much of our digital lives happen. But taking advantage of that position means solving a problem that's as much about trust and product philosophy as it is about AI.

The technology that made the old Siri feel like it was constantly forgetting you is disappearing. Now Apple has to decide how much it wants the new Siri to remember.

More from Tom's Guide

• Claude Sonnet 5.5 just launched — here's why it's about to become your daily driver
• I asked ChatGPT, Gemini, Claude and Muse to give me new podcasts based on my favorites — these 7 picks got added to my queue for International Podcast Day
• 'Surveillance Tamagotchi': Tom's Guide readers weigh in on Meta's new Muse Charm

OpenAI's Dots essentially give ChatGPT a job rather than another prompt. Each Dot gets its own computer and browser in the cloud, can connect to the apps you already use and can keep working toward your goals even when you're no longer sitting in front of ChatGPT.

As someone who spends a lot of time testing AI agents, I immediately wanted to see what one could actually do.

There's just one problem: I can't use Dots on my ChatGPT Plus account. And neither can most ChatGPT users.

OpenAI is initially rolling out Dots to ChatGPT Pro and Business Premium users in eligible markets. Enterprise users, including those on Edu and Healthcare plans, can also try the beta if their workspace administrator enables it.

OpenAI says it plans to expand Dots to more users "soon," but there's no date for when Plus, Go or Free users might get access.

At first, keeping one of DevDay's biggest announcements away from most ChatGPT users seems like an odd decision. But the more you look at what Dots are capable of — and particularly the safeguards OpenAI has built around them — the more the cautious rollout starts to make sense.

Dots aren't just another version of ChatGPT

The important difference between Dots and the ChatGPT most of us use today isn't how smart the AI is. It's what happens when you stop talking to it. ChatGPT generally waits for you to tell it what to do next. A Dot is designed to keep going.

Powered by GPT-6 Astra, Dots can work on several projects at once, learn from your feedback and continue working toward your goals 24/7. OpenAI says they can connect to more than 4,000 apps through its plugin ecosystem, and you can reach your Dot through ChatGPT, Slack or Teams.

More importantly, a Dot can look for useful work even when you haven't explicitly given it another prompt.

OpenAI calls this "proactive research." One early tester's Dot, for example, noticed that he'd forgotten to invoice a publication, prepared the invoice and sent it after receiving his approval.

Instead of opening ChatGPT because you need something, OpenAI wants Dots to notice that something needs doing and bring it to you. It's also where things get considerably more complicated.

Giving AI more freedom creates a bigger problem

If ChatGPT misunderstands a prompt, you might get a bad answer. If an AI agent misunderstands what you wanted while it has access to your apps, browser and potentially your computer, the stakes are higher.

OpenAI appears well aware of that, because Dots come with a surprisingly extensive set of restrictions around what they can do without you.

When a Dot is performing proactive research in the background, for example, its connected-app tools are restricted to read-only access. That means it can't independently send messages, modify content inside an app or control your browser or computer while looking for something useful to do.

You also choose which apps a Dot can access, while Custom Rules can determine which actions it can take independently, which require your approval and which are blocked.

Then there's auto-review. Actions that could affect your accounts or share information are checked against your instructions and OpenAI's safety requirements before they're allowed to proceed. Some particularly sensitive jobs, such as changing a password, always stay with you.

As you can immediately see, these aren't the kinds of guardrails you need when you're asking ChatGPT to write an email or plan dinner. They are what you need when an AI can actually do things.

OpenAI's own testing shows how difficult that can get

OpenAI's safety testing also gives us a clue as to why Dots may be getting a more limited rollout. One of the challenges with an always-on agent is keeping track of exactly what you've authorized it to do as it moves between tasks.

Giving an AI permission to take an action in one situation doesn't necessarily mean you want it taking the same action later. And the more jobs an agent is juggling, the more complicated those boundaries become.

That's particularly important with GPT-6 Astra, the model powering Dots. OpenAI's broader Astra testing found that the model was considerably better at staying within its authorized scope than previous models, but it wasn't perfect. In simulated work environments using the confirmation policy that users receive by default, Astra still produced a 3% overall rate of outcomes OpenAI classified as misaligned. That was nevertheless substantially lower than the 8% recorded by GPT-5.6 Sol under the same conditions.

That doesn't mean your Dot has a 3% chance of going rogue every time you give it something to do. These are deliberately difficult evaluations designed to expose potential problems, not a prediction of how often something will go wrong in everyday use. But they demonstrate the problem OpenAI is trying to solve.

Building an AI capable of doing more on its own is one thing. Making sure it consistently understands where its authority ends is another.

Dots may also be expensive to run

There's another potential reason OpenAI isn't immediately handing Dots to every ChatGPT user: all that work requires computing power. OpenAI hasn't said cost is why Dots aren't available on Plus, Go or Free, so we can't know for sure whether that was part of the decision.

But the way Dots are being packaged is interesting. Pro and Business Premium subscribers get their first Dot as part of their existing plan, and OpenAI says conversations with it won't count toward normal ChatGPT usage limits. The plan also comes with an allowance for what OpenAI calls "deeper work."

Eventually, OpenAI says users will be able to add more Dots and increase either how quickly a Dot works or the total amount of work it can perform each month.

That's a pretty strong indication that there's a meaningful cost attached to having an AI agent working for you around the clock, even if OpenAI isn't saying that's why access is limited today.

And it raises an interesting question about what Dots might eventually look like on cheaper ChatGPT plans.

A future Plus version could conceivably come with tighter limits on background work, fewer resources or additional charges once you hit a certain amount of usage. For now, though, OpenAI hasn't announced how access will work beyond saying more users are coming "soon."

What can you use instead of Dots?

The good news is that you don't necessarily have to wait for Dots to start handing some of your work over to ChatGPT.

The closest option is ChatGPT Work. While it isn't an always-on personal assistant in quite the same way as a Dot, Work can take a larger job and carry it through multiple steps for you, including researching information, working across files and apps and navigating web-based tasks.

That's useful when you want to hand ChatGPT a job rather than sit there prompting it through every individual step.

Scheduled tasks cover another piece of the Dots experience. You can have ChatGPT run something later or on a recurring schedule, which works well for things like regular briefings, reminders or checking something periodically.

Neither gives you the full experience OpenAI is promising with Dots, but it's a start.

Meta's Muse is another option if the always-on assistant part is what interests you most. Muse takes a similarly agentic approach and offers a useful glimpse at where this category of AI assistant is heading.

There just isn't a true one-for-one replacement for Dots yet.

Bottom line: I'm waiting for Dots to come to Plus

For now, OpenAI hasn't given us a date for when that will happen. The company says only that it is beginning with Pro, Business Premium and Enterprise and plans to expand Dots to more users soon.

Still, the limited rollout may be one of the most interesting things about Dots. Typically, most major ChatGPT upgrades have been about giving us a smarter model, a new tool or another way to ask AI for something. But Dots changes that completely. Now, OpenAI is building an AI that can continue working when you're gone, notice things you haven't asked about yet and potentially take action on your behalf.

So, while I'd like a Dot on my Plus account right now, I can understand why OpenAI isn't handing one to everyone on day one. The real test will come when it does.

More from Tom's Guide

• Claude Sonnet 5.5 just launched — here's why it's about to become your daily driver
• I asked ChatGPT, Gemini, Claude and Muse to give me new podcasts based on my favorites — these 7 picks got added to my queue for International Podcast Day
• 'Surveillance Tamagotchi': Tom's Guide readers weigh in on Meta's new Muse Charm

For many ChatGPT users, the AI is mostly a tool to answer questions in a chat window. You type a prompt, wait for an answer and then take that answer somewhere else yourself. OpenAI is trying to change that with Dots.

Announced today, Dots are persistent AI agents that get their own virtual computer in the cloud, allowing them to browse the web, use connected apps and continue working even when you're no longer sitting in front of ChatGPT.

Instead of asking ChatGPT to help you accomplish something and then doing the rest yourself, the idea is that you can hand a Dot the job.

That could mean comparing how much dinner would cost from several restaurants, keeping tabs on something you're researching or catching an administrative task you've forgotten about. And because the Dot keeps running in the background, you don't necessarily need to keep ChatGPT open while it works.

It's a significant change in how OpenAI wants us to use ChatGPT. Rather than simply making the chatbot smarter, Dots are designed to give it something it has largely been missing: a place to actually do things for you.

ChatGPT gets its own computer

The easiest way to understand a Dot is to imagine giving ChatGPT its own computer.

Each Dot operates inside an isolated virtual machine running in the cloud rather than taking over your laptop or filling your browser with tabs. From there, it can browse websites, open applications and carry out tasks using the services you've connected to it.

You can explicitly give a Dot permission to assist with something on your local machine, but otherwise the two remain separate.

Dots can also connect to more than 4,000 apps through plugins, giving the agent a way to work with services you already use rather than keeping everything contained inside ChatGPT.

And you don't have to constantly return to the ChatGPT website to communicate with it.

OpenAI says you'll be able to reach a Dot through voice calls and services including Slack and Microsoft Teams. SMS support is also planned, meaning you could eventually text your AI assistant much like you'd text another person.

That starts to make the "persistent" part of Dots considerably more interesting. ChatGPT doesn't just exist in the tab where you started the conversation anymore.

What would you actually use a Dot for?

The best examples aren't complicated autonomous coding jobs or elaborate demonstrations of AI agents. They're the annoying everyday tasks that currently require you to bounce between several websites or apps.

Take ordering dinner. You could ask a Dot to look at several options for dinner, and instead of simply recommending restaurants, it could browse delivery services and build itemized cart estimates for them. It can factor in things such as estimated local taxes and a tip allowance and then send you the comparisons so you can decide which one you actually want to order.

That's a fundamentally different interaction from asking ChatGPT, "Where should I order dinner tonight?"

The same applies to work. Give a Dot an interview recording or video transcript and it could help produce social captions and show notes, identify potential clips and work with assets across connected platforms.

Or it might notice an administrative job you've forgotten about, such as an invoice you still need to send, prepare it and put it in front of you for approval. Once you start, you'll discover the difference is how much of the process happens after you've made the initial request.

Until now, even when ChatGPT gave me exactly what I wanted, I was usually the one copying the answer, opening another app and doing something with it. Dots are an attempt to remove more of those steps.

You can talk to your Dot almost anywhere

One of the more interesting parts of the launch is that OpenAI doesn't expect you to spend all day watching your Dot work. You'll initially need to create one using ChatGPT on the desktop web or desktop app, where you can choose its purpose and connect the services it needs.

After that, the relationship becomes less tied to the ChatGPT interface. You can message your Dot from mobile and, depending on the service, communicate with it through other platforms. Voice calls provide another way to check in or give it something new to do, while upcoming SMS support could make interacting with one feel even less like using a conventional chatbot.

We've already seen AI companies add intelligent agents to chatbots, but most still expect you to go to the AI. Dots begin flipping that relationship around by allowing the AI to exist in more of the places where you're already communicating.

How much do Dots cost?

Dots are rolling out today to ChatGPT Pro and Business Premium subscribers in eligible markets. OpenAI is also making them available through Enterprise betas where administrators choose to enable the feature.

For Pro subscribers, the first Dot is included with the subscription at no additional charge, and OpenAI says users will receive extended usage limits during the first month.

You'll need to create and initially configure a Dot through ChatGPT on desktop, including connecting any apps you want it to use. Once that's done, you'll be able to communicate with it from mobile as well.

That makes Dots one of the more substantial additions OpenAI has made to its paid ChatGPT experience, particularly for Pro users who already have access to its most capable models and agent features.

ChatGPT is starting to leave the chat window

I'll need more time actually using Dots before deciding whether having an always-on AI assistant is enough to justify the cost of ChatGPT Pro on its own. But I'm more interested in what Dots say about where ChatGPT is going.

For the past few years, the basic relationship hasn't changed all that much even as the models behind ChatGPT have become dramatically more capable. I open ChatGPT, ask it for something and wait for it to respond.

Dots start breaking that pattern. If they work as OpenAI intends, I could give ChatGPT something to do in the morning, close the tab and spend the next few hours doing something completely unrelated while its virtual computer keeps working.

Then I could text or call it, see what it found and decide whether I want it to take the next step.

The takeaway: exciting or scary

If you're like me and think giving ChatGPT its own computer sounds a little terrifying, there are certainly pretty obvious questions about what happens when the AI gets something wrong. OpenAI has built several restrictions into Dots that are particularly important when the agent is operating in the background.

For starters, your Dot's cloud computer is isolated from your personal computer by default. Your local files aren't suddenly sitting inside the virtual machine just because you've created a Dot.

There's also a separation between the AI model and your credentials. Dots can sign into supported services using saved credentials without exposing the password itself to the underlying model.

More importantly, OpenAI isn't giving Dots unlimited permission to quietly act on your behalf. For example, during proactive background research, Dots operate in read-only mode. That means an agent can gather information while you're away, but it can't independently send a message, delete a file or spend your money.

Actions like those require you to come back into the loop and approve them. That's important, because an AI capable of researching dinner prices overnight is very different from one capable of ordering $200 worth of sushi because it misunderstood what you wanted.

It doesn't eliminate the privacy and security questions surrounding persistent agents, particularly as people connect more accounts to them, but it puts some boundaries around what "always on" actually means.

Will you try Dots? Let me know your initial reactions in the comments.

More from Tom's Guide

• Claude Sonnet 5.5 just launched — here's why it's about to become your daily driver
• I asked ChatGPT, Gemini, Claude and Muse to give me new podcasts based on my favorites — these 7 picks got added to my queue for International Podcast Day
• 'Surveillance Tamagotchi': Tom's Guide readers weigh in on Meta's new Muse Charm

Meta has a new AI model, and I wanted to know how it handles the kinds of prompts people actually use, especially in comparison to ChatGPT-6.

Meta Muse is the company’s latest model, arriving at a time when ChatGPT has become the default chatbot for millions of people. But benchmarks only tell us so much. What matters more to me is whether an AI can write the awkward text you’ve been avoiding, plan a realistic day out with your family or explain something complicated without making it even more confusing.

So I put Meta Muse and ChatGPT through five everyday prompts covering writing, planning, problem-solving, explaining and creative reasoning. I wanted to know which one best understood what I was actually asking for and which could give me the best answer to use

After five rounds, there was a clear winner — but the results were much closer than I expected.

1. Handling an awkward situation

Prompt: My neighbor keeps putting their trash cans partly in front of my driveway, and it's making it difficult to back out. Write a text asking them to stop. I want to be friendly because we get along, but clear enough that they actually change what they're doing. Don't make it sound passive-aggressive.

ChatGPT led directly with the request, cutting to the chase and delivered a note that eliminated ambiguity around how far over they need to go.

Meta delivered a courteous and natural note. The ask feels neighborly and the phrasing suggests a minor
physical adjustment rather than something more serious.

Winner: Meta wins for a draft that hits the prompt’s tone constraints better for an ongoing, friendly relationship. ChatGPT’s message of “completely clear” reads like an HOA notice and slightly passive-aggressive.

2.  Planning under constraints

Prompt: Plan a Saturday in New York City for a family with three kids ages 5, 9 and 12. We have a budget of $200 for activities and food, want to avoid spending the whole day walking, and need at least one indoor activity in case it rains. Give me a realistic schedule from 10 a.m. to 7 p.m.

ChatGPT did a nice job with geographic clustering, to ensure the walking distance was short for the first part of the day. The zero-cost options and pacing of the day with high energy and sit-down segments were also helpful.

Meta picked an attraction with far higher entertainment value for the targeted age spread (5, 9, and 12). The American Museum of Natural History (AMNH) easily captivates all three ages simultaneously. Meta also demonstrated accurate NYC transit knowledge by explicitly citing the MTA’s 44-inch height cutoff for free child rides, included a practical playground stop right across the street, and gave concrete logistical guidance (e.g., sitting on the right side of the ferry outbound).

Winner: ChatGPT wins. While Meta's AMNH plan is more fun on paper, its entire budget math hinges on an unstated assumption of tri-state residency. ChatGPT wins on strict adherence to the prompt's financial and logistical constraints.

3. Advice / problem-solving

Prompt: I have 45 minutes before guests arrive. My kitchen is messy, the living room needs vacuuming, there's laundry on the couch and I still need to shower. Give me the most efficient plan to get everything done. Tell me what I should skip if I run out of time.

ChatGPT anticipated a critical blind spot the prompt forgot to mention: the guest bathroom. Adding a 5-minute block to check toilet paper and wipe the sink is the difference between a good plan and a realistic one. It also provided highly specific guardrails for the shower block, which prevents the most common reason people run late in this scenario.

Meta delivered behavioral advice, which was excellent, and explicitly stated the rationale for the timeline and ending with the mandate to use a timer. It’s timing format was also easier to read than ChatGPT’s.

Winner: ChatGPT wins for operational realism and anticipating unstated needs. While Meta's timer advice is sharp, its kitchen expectations are a setup for failure. ChatGPT recognized that when you have 45 minutes, you are mostly straightening and organizing, not necessarily cleaning it in such a short amount of time.

4.  Explain something complicated simply

Prompt: Explain why airplanes can stay in the air to a curious 10-year-old. Don't use equations or assume they know what lift, drag or air pressure mean. Keep it under 200 words and use an analogy they could picture.

ChatGPT introduced an accurate, intuitive physical mechanism: wings push air down, which pushes the plane up (Newton’s third law translated into plain English).

Meta offered perfect audience calibration. The tone is lively and speaks directly to a 10-year-old ("with a grown-up driving, of course!", "whoosh"). It picked one analogy—the hand out the car window—and committed to it entirely, tying it back to paper airplanes and the runway to explain why forward speed matters.

Winner: Meta wins. It kept the energy high, matched a 10-year-old's reading level without talking down to them and delivered a clear, memorable takeaway.

5. Creative reasoning

Prompt: Invent a new app that solves a small but annoying everyday problem people have at home. It can't be a to-do list, meal planner, budgeting app or smart-home remote. Give it a name, explain the problem it solves, describe how it works and tell me why someone would actually keep using it after the first week.

ChatGPT solved a universal human frustration without requiring whole-house cataloging. The interaction model involves snapping a photo and speaking natural language ("Christmas lights in the blue bin") is low-effort.

Meta identified a hyper-specific, universally relatable domestic failure point and gave the app a brilliant, punchy name (Takeout). Its retention logic is rock-solid: instead of pretending users love opening chore apps, it acknowledged that the best utility apps require zero maintenance and rely on loss aversion.

Winner: Meta wins because the chatbot built a product around the reality of household laziness and recognized that an app survives past week one only when it demands almost no manual input from the user.

Final thoughts

After five prompts, Meta Muse takes the win 3-2. To say I'm both shocked and impressed is a serious understatement. While ChatGPT was stronger when a prompt required careful planning and attention to practical constraints, Meta Muse stood out when the task depended more on understanding people. Its neighbor text sounded like something I'd actually send, its airplane explanation was perfectly pitched for a 10-year-old and its app idea showed a surprisingly good understanding of why people abandon apps that require too much effort.

Meta Muse feels very much like it's own model, not one that's trying to imitate or catch up to ChatGPT. The little "Muse" at the top is unnecessary, but adorable.

ChatGPT is still my go-to, but this test proves that Meta has come a long way and this latest model offers a pretty strong first impression.

More from Tom's Guide

• Nvidia says 'local AI is here' — and it could change how you use AI at home
• OpenAI is paying influencers to convince you to use ChatGPT — here’s what we know
• I tested ChatGPT-6 vs Claude Opus 5.5 with 5 everyday prompts — it wasn’t even close

OpenAI and Anthropic both released new flagship AI models this month, ChatGPT-6 and Claude Opus 5.5. Both are ridiculously capable to the point that they do so much more than what most users prompt them to.

OpenAI's GPT-6 family builds on GPT-6 Astra, the company's most impressive model yet, with major gains in areas including reasoning, professional work, coding, browsing and computer use. The newer GPT-6 Sol brings much of that intelligence into the version designed for everyday work, with OpenAI positioning it as a faster and more affordable reasoning model.

Claude Opus 5.5 comes at the problem from a slightly different direction. Anthropic describes it as a model built for long-running agentic and knowledge work, with adaptive thinking that determines how much effort a request needs. It has a 1 million-token context window, can produce up to 128,000 output tokens and, according to Anthropic, reaches roughly the same performance as its higher-end Claude Fable 5.1 on most tasks while costing 40% less to run than the previous Opus. Anthropic has also specifically highlighted improvements to how naturally the model communicates and follows writing instructions.

Those differences sound impressive, but they don't necessarily tell you which chatbot you'd rather have helping you in real life.

So I gave ChatGPT-6 and Claude Opus 5.5 the exact same five prompts. The result wasn't the clean sweep I expected. Here's what happened when I put them through every day prompts.

1.  Messy real-world planning — constraints + judgment

Prompt: I have $150 to plan a birthday party for 10 kids ages 8–11. It needs to last three hours, work indoors if it rains, include food, and keep screen time to a minimum. Two kids are vegetarian and one has a peanut allergy. Create the full plan, including a budget, timeline, activities and backup plan. Do not go over $150.

ChatGPT offered clear conceptual framing by giving the event an immediate identity and direction. Budgeting $36 for 3 delivery cheese pizzas is practically easier for a parent hosting 10 kids than managing DIY mini-pizzas in a standard home kitchen. It also recognized the need for an "Emergency/price buffer" at the end.

Claude was smart for being proactive about allergies and dietary protocols. It also addressed cross-contamination of bakery cakes and boxed mixes. It also did double-duty budgeting by spending 14 on canvas tote bags and $12 on fabric markers serves as both an arrival craft activity and the take-home party favor, saving money and cutting down on cheap plastic filler. It also offered an actionable, timed schedule and operational troubleshooting.

Winner: Claude wins because it delivered a turnkey, host-ready plan with genuine safety diligence, whereas ChatGPT produced a fragmented grocery list.

2. Writing — make AI sound human

Prompt: Rewrite this boring announcement so it sounds like something a real person would actually want to read. Keep every factual detail, don't use em dashes, avoid clichés, and don't use the words “exciting,” “revolutionary,” “game-changing” or “seamless.”

“Our company is introducing a new AI-powered calendar feature that automatically identifies scheduling conflicts, recommends meeting times and creates summaries after meetings. The feature launches October 15 and will be available to all customers at no additional cost.”

ChatGPT offered a punchy one-liner that draws the reader in and it also avoided AI styles and retained every factual detail. The paragraph breaks make it easy to scan quickly.

Claude reframed the bullet points into concrete human benefits and reframed the details around actual workplace relief.

Winner: Claude wins because it understood that a "human" rewrite requires rethinking how to explain the value of the features, rather than just shifting a few verbs in the original text.

3. Reasoning — the answer isn't obvious

Prompt: A family is choosing between two vacation rentals. Rental A costs $1,850 and is a 10-minute walk from the beach. Rental B costs $1,500 but requires a 25-minute drive to the beach and $35 per day for parking. They're staying seven days and expect to visit the beach twice a day. They have three kids. Which should they choose? Don't just calculate the cost. Identify the factors that could change your recommendation and tell me what additional information would matter most.

ChatGPT provided the single best counter-perspective on kids' ages. It also smartly pointed out that a "10-minute walk" varies wildly depending on infrastructure (boardwalk vs. crossing a four-lane highway or traversing sand dunes). The chatbot specifically noted that Rental A allows parents to split up (e.g., one parent takes a tired child back while the other stays at the beach with the other two).

Claude actually did the mileage math, proving that the financial savings of Rental B is effectively a negligible $55–$65 total for the entire week. Claude also made an astute psychological point: loading/unloading three kids into car seats four times a day means the family will likely abandon the second daily beach trip entirely. It also correctly flagged vacation rental platforms' dirty secret: base rates don't reflect cleaning fees, service fees, or occupancy taxes, which could easily flip the $105 difference on their own.

Winner: Claude wins by a hair. Although ChatGPT had the best individual point (realizing that walking with heavy beach gear and young kids can actually be more grueling than driving), Claude won overall because its financial breakdown was thorough down to the gallon, it anticipated hidden fees, and its observation that the commute would cause the family to cancel their planned outings was spot-on.

4.  Option overload — find what's important

Prompt: A city has $10 million to spend and three options:

A) Build a new public library expected to serve 200,000 visits per year for at least 30 years.
B) Give every household a one-time $500 payment.
C) Upgrade aging water infrastructure, which residents won't notice unless something goes wrong.

The city can only choose one. Choose the option you think is best. Explain your reasoning, identify the strongest argument against your choice, and explain what additional piece of information would be most likely to change your answer. Do not hedge by saying all three are good options.

ChatGPT followed all instructions cleanly and delivered an easy-to-read defense of option C. It also cleanly defined the engineering assessment threshold.

Claude framed the choice it made through municipal incentives and recognized that funding what electoral politics ignores is the core function of city management.

Winner: Claude wins for answering significantly better. The chatbot seemed liked it had experience in public-finance administration whereas ChatGPT’s response felt more standard and surface-level.

5  The killer test — vague request + initiative

Prompt: My life feels disorganized. I have three kids, a full-time job, household responsibilities and about 30 minutes every evening to get things under control. Build me a system I could realistically maintain. Don't give me generic productivity advice. Make reasonable assumptions where necessary, tell me what you're assuming, and give me something I can start tonight.

ChatGPT offered simple ways to prevent burnout and included instructions that were broken down into steps so the user doesn’t have to plan or try to guess where to start.

Claude addressed the mental load, which is a sharp, accurate insight into why disorganization happens. It also mentioned a useful tip of delegating when possible chores while recognizing that one parent cannot be the sole operational engine for a family of five.

Winner: ChatGPT wins because its cognitive framework is much more protective of an exhausted parent’s mental bandwidth. Claude gave a good domestic organizing routine; ChatGPT gave a decision-making filter that actually keeps work, family, and life from cannibalizing each other.

Overall winner: Claude

After five tests, Claude Opus 5.5 came out ahead in four. But what surprised me was how consistently the same difference between these models showed up across completely unrelated tasks. Claude tends to keep digging.

When I asked it to plan a birthday party, it thought about cross-contamination and found ways for one purchase to serve two purposes. When I gave it the vacation rental problem, it calculated costs I hadn't explicitly asked for and considered whether the inconvenience of driving would actually change the family's behavior. Even when choosing how a city should spend $10 million, it looked beyond the obvious pros and cons to think about why governments make these decisions in the first place.

That's where Opus 5.5 felt strongest. ChatGPT-6 was generally more streamlined, which is very much still on-brand for ChatGPT. Regardless of the model, ChatGPT always seems to be the less "fluffy" of the two models, with a response that gets to the point faster. That also means that, at times, its answers didn't go quite as deep as Claude's. However, that approach became an advantage in my final test.

More from Tom's Guide

• Nvidia says 'local AI is here' — and it could change how you use AI at home
• SpaceX is launching Google’s next big AI data center hundreds of miles into space
• AI could be costing you money: new study finds chatbots get most financial questions wrong

If it feels like ChatGPT is showing up in more places than your browser or phone lately, it's no coincidence.

OpenAI is reportedly ramping up its use of influencers to promote ChatGPT, and it isn't limiting those partnerships to the usual tech creators talking about prompts, models and benchmarks.

Instead, the company is working with creators in areas including parenting, fitness, education, travel, shopping and careers, according to a new report from Business Insider. The goal appears to be less about explaining what ChatGPT is and more about showing how the chatbot can fit into everyday life.

And the number of sponsored posts is climbing quickly. According to HypeAuditor data cited by Business Insider, Instagram posts using #chatgpt_partner or #chatgptpartner jumped from 61 in June to 122 in July and 141 in August.

That's more than double the number of sponsored posts in just two months, and it offers a glimpse at how OpenAI is trying to turn ChatGPT into something you reach for throughout your day.

OpenAI wants ChatGPT to be more than your favorite chatbot

What's particularly interesting about OpenAI's influencer strategy is the type of content it's promoting. Business Insider reports that OpenAI has been interested in working with fitness creators, for example, because people already use ChatGPT for questions about workouts and running.

A sponsored back-to-school campaign showed a parent using ChatGPT to help her children with math homework while highlighting features including Study Mode and teen-safety tools.

Meanwhile, The Shade Room, which has more than 28 million Instagram followers, published a sponsored ChatGPT word-search puzzle. All of this highlights how OpenAI is working to encourage ChatGPT name recognition along with everyday use cases.

OpenAI now appears to be marketing individual reasons to keep using it, whether you're planning a workout, helping with homework or trying to get something done at work.

One creator who has worked with OpenAI told Business Insider that her audience already has ChatGPT. The challenge, then, isn't persuading people to try AI for the first time. It's getting them to choose ChatGPT when they do.

OpenAI is building a creator operation

There are signs this isn't just a handful of curated sponsorships. OpenAI has been expanding its creator partnerships operation and has sought employees who can help the company work with creators and influencers at scale.

Earlier this year, OpenAI also hired Charles Porch, who spent years overseeing partnerships at Instagram, to take on the role of Vice President of Global Creative Partnerships for OpenAI , according to his Linkedin profile. The company has even experimented with the kind of influencer trips normally associated with beauty, fashion and lifestyle brands.

In August, OpenAI hosted creators at a luxury retreat in New York where attendees participated in activities and workshops while learning about OpenAI products. The trip also demonstrated one of the risks of this strategy.

Some participating influencers faced criticism from followers over issues including AI's environmental impact and whether creators were compromising their independence by partnering with OpenAI. TechCrunch reported that at least one creator appeared to remove a video about the event following the backlash.

How to tell when a ChatGPT post is sponsored

There's nothing inherently unusual about OpenAI paying influencers to promote ChatGPT. Brands have used creator partnerships for years. But it's worth knowing when the enthusiastic ChatGPT recommendation appearing in your feed is an independent recommendation and when it's advertising.

The Federal Trade Commission says creators should clearly disclose when they have a material relationship with a company they're endorsing. That can include being paid, receiving free products or getting other perks.

The disclosure should also be difficult to miss rather than buried among hashtags or hidden behind a "more" button.

In OpenAI's case, hashtags including #chatgpt_partner and #chatgptpartner are one obvious clue that you're looking at sponsored content.

So, the next time a creator casually shows how ChatGPT helped plan a vacation, explain homework or build a workout, it's worth checking the caption. You may be watching an artfully curated ad campaign.

More from Tom's Guide

• Nvidia says 'local AI is here' — and it could change how you use AI at home
• I replaced ChatGPT with Siri AI for a day — here's what surprised me
• AI could be costing you money: new study finds chatbots get most financial questions wrong

There’s something deeply ironic about getting fired from an AI company for using AI, but that’s reportedly exactly what happened to multiple contractors helping train OpenAI’s models.

Contractors hired to review and improve the responses generated by OpenAI’s AI models have been fired or removed from projects after using AI to do their own work, according to a new report from 404 Media.

The rule makes more sense than it might sound. These workers are being hired specifically for their human judgment — something that becomes considerably less useful if ChatGPT or another AI chatbot is doing the judging for them.

One contractor told 404 Media that they see people using AI “all the time” and that workers are regularly removed for doing it. Another contractor shared what they said was a termination letter citing problems with the “authenticity” of their work.

Keeping humans in the loop

AI models don't simply learn everything from scraping enormous amounts of text from books or the internet.

OpenAI says information provided or generated by human trainers and researchers is one of the sources used to develop its foundation models. Human feedback has also long played a role in teaching models which responses people prefer and how those models should behave.

That makes the quality of the human feedback truly important. OpenAI has previously acknowledged in its research that models trained using human feedback can be influenced by the people labeling that data.

The contractors described by 404 Media work across various OpenAI projects, some of which can involve thousands of workers. Their jobs can include reading AI responses, rating them and deciding which answers are better.

But according to internal documents obtained by 404 Media, reviewers aren't allowed to outsource that judgment back to an AI. One set of instructions reportedly prohibits workers from using AI to write feedback or comments and specifically names tools including Grammarly and AI translation.

Mercor, an AI-training company that employs contractors working on OpenAI projects, confirmed to 404 Media that its contracts prohibit workers from using large language models to complete projects.

AI detectors didn't catch the cheating

You might assume an AI company would catch workers using AI with an AI detector. Apparently not.

404 Media doesn't reveal exactly how the contractors who were fired were caught. But internal instructions obtained by the publication offer a fascinating look at how reviewers are told to spot suspected AI use.

And AI detectors aren't part of the process.

The instructions reportedly tell reviewers not to use tools such as GPTZero because they aren't considered reliable. Instead, reviewers are told to judge the overall pattern of a worker's submissions and look for clues including repetitive wording, unusually fast completion times and writing patterns associated with AI. Even excessive use of em dashes can raise suspicion.

Reviewers are also reportedly instructed not to tell workers exactly what made them suspect AI use, because doing so could make it easier to hide in the future.

This is surprisingly revealing about the process: the people checking whether humans are secretly using AI to help train ChatGPT are explicitly told not to trust AI-detection software to do it for them.

Bottom line: humans are irreplaceable

OpenAI uses a lot of synthetic data — information generated with the help of AI models — as part of some training processes. The company says synthetic data can improve model performance and fill gaps where other training data is scarce.

But that's different from secretly substituting an AI model for a person who has specifically been hired to provide human feedback. If an evaluator is supposed to tell OpenAI which response a human prefers, asking another AI to make that decision defeats much of the point of hiring the evaluator.

It doesn't mean these incidents have made ChatGPT worse, and there's no evidence in 404 Media's reporting that AI-generated contractor work caused measurable damage to an OpenAI model.

But the situation does expose one of the stranger realities of the AI boom: even the companies building the AI systems still need humans to tell those systems when they've gotten something right.

More from Tom's Guide

• Nvidia says 'local AI is here' — and it could change how you use AI at home
• I replaced ChatGPT with Siri AI for a day — here's what surprised me
• AI could be costing you money: new study finds chatbots get most financial questions wrong

An independent researcher has uncovered what appears to be a cross-site tracking system connected to OpenAI's rapidly growing advertising business. At the center of it is a cookie called "__obi," which can reportedly travel from ChatGPT to websites using OpenAI's advertising technology.

That means if you use ChatGPT and later visit a participating retailer or other website, OpenAI may be able to recognize the same browser and receive information about what you do there.

The finding comes at a particularly interesting time. OpenAI has been rapidly expanding advertising inside ChatGPT, recently announcing new AI-powered advertising tools and Sponsored Agents. The company said last month that ChatGPT Ads had already reached a $1 billion annualized revenue run rate.

But this isn't quite as simple as assuming ChatGPT is watching everything you do online. There are some important limitations to the research, including one major piece of the system the researcher wasn't able to independently verify.

Here's what we know.

What is __obi?

The research comes from an independent researcher who publishes under the name Buchodi and says they analyzed months of network traffic involving 936 OpenAI advertiser pixels across 1,029 hostnames.

According to the analysis, visiting ChatGPT can result in OpenAI placing a cookie called __obi in your browser.

Cookies themselves aren't unusual; you've probably accepted them on a several websites. They're small pieces of information stored by your browser and are used for everything from keeping you logged into a website to remembering preferences and measuring advertising.

What's interesting here is how this particular cookie is configured. Buchodi found that __obi can last for one year and uses the SameSite=None setting, which allows it to be included in certain requests originating from other websites.

That's where OpenAI's advertising system comes in. Businesses advertising through ChatGPT can install OpenAI's measurement pixel on their websites, similar to the tracking technology used by advertising platforms such as Google and Meta.

When Buchodi visited participating websites, requests sent back to OpenAI included the __obi identifier. In other words, the same identifier created through ChatGPT could show up again when that browser visited another website.

What can OpenAI see?

Potentially more than simply whether you visited a website. According to Buchodi's network analysis, OpenAI's advertising collector received information associated with pages people visited and conversion events on participating sites.

The researcher examined 23,929 observed URLs and found that query strings were stripped before being sent to OpenAI, but URL paths remained.

A URL path can sometimes reveal considerable information about what someone is looking at. Buchodi says paths observed in the dataset included pages involving a medical condition, debt services and a litigation intake form.

The advertising system also supports what OpenAI calls conversion measurement.

OpenAI's own documentation says advertisers can install a pixel and use conversion measurement to understand what happens after someone interacts with a ChatGPT ad. Its measurement integrations can support conversion tracking and attribution across websites, apps and even offline activity.

That's standard practice in digital advertising. What's unusual is that the advertising platform in this case sits alongside an AI assistant where users may share substantially more personal context than they would with a traditional search engine or social network.

Is it connected to your ChatGPT account?

This is where the story gets more complicated. Buchodi found that the process used to generate the identifier included tokens labeled either account_user or anonymous.

Of 932 tokens the researcher decoded, 736 were associated with logged-in users and 196 were labeled anonymous. The researcher also found that anonymous identifiers could remain stable on the same device for at least 27 days.

However, there's an important caveat. Buchodi did not directly observe OpenAI taking an event received from an advertiser website and joining it to a person's ChatGPT account on its servers.

The researcher argues that the architecture makes such a connection possible, but that final server-side step wasn't independently observed.

So it would be premature to say OpenAI definitively attaches every page you visit on participating sites to your ChatGPT profile. What the research does demonstrate is that the same OpenAI identifier can be transmitted from participating websites back to OpenAI.

Buchodi says the identifier appeared across a number of recognizable websites during testing. On one device, the same __obi value was sent to OpenAI from 12 commercial websites using 13 separate advertising pixels. Across the larger dataset, 12 of 30 identifiers appeared with more than one advertiser, while one appeared across 10 different advertisers.

The researcher also notes several important limitations. Testing was performed using Chrome on Android, and desktop Chrome wasn't tested. Safari's anti-tracking protections block the third-party cookie mechanism described in the research, meaning it shouldn't operate the same way on Safari or other iOS browsers using Apple's WebKit.

The identifier also wasn't generated during every ChatGPT session. Buchodi estimates that roughly one in five sessions tested produced the synchronization token.

What does OpenAI say?

OpenAI hasn't publicly responded to the specific findings. Buchodi says they contacted OpenAI's press and privacy teams on September 14 and asked why __obi was classified as an analytics cookie and how it behaves when someone rejects marketing cookies.

According to the researcher, OpenAI acknowledged the inquiry and said the findings would be passed along internally but had not provided answers to those questions when the research was published.

Tom's Guide has contacted OpenAI for comment and will update this story if we hear back.

OpenAI's public documentation emphasizes that advertisers do not receive ChatGPT conversations, chat histories, memories or personal details.

The company also says it does not sell user data to advertisers, and advertisers receive aggregated reporting about things such as views, clicks and conversions.

That's an important distinction here: the research does not suggest advertisers themselves can read your ChatGPT conversations or access the __obi identifier.

Instead, the concern is about information potentially flowing in the other direction — from websites using OpenAI's advertising tools back to OpenAI.

Why this matters

Cross-site advertising measurement isn't new. In fact, Google, Meta and other advertising companies have spent years building systems designed to determine whether someone who saw or clicked an ad later visited a website or bought something. But ChatGPT introduces another dimension.

People don't just search for products using ChatGPT. They use it to discuss relationships, health concerns, finances, careers, travel plans and countless other parts of their lives.

OpenAI itself promotes ChatGPT advertising around this idea, telling advertisers that people share "richer context" in ChatGPT while exploring options and making decisions.

That's why the boundary between your ChatGPT activity and what you do elsewhere on the web matters.

For now, the evidence doesn't show that OpenAI is handing your conversations to advertisers — and OpenAI explicitly says it doesn't.

But the discovery of __obi raises a different question: how much information should be able to flow back into an AI platform after you leave it?

What do you think? Share your thoughts in the comments.

More from Tom's Guide

• AI is turning ordinary photos of kids into sexual abuse imagery — here's how to protect your family
• OpenAI paid contractors to read ChatGPT conversations — here's how to protect yourself
• I stopped reading the fine print myself after discovering this overlooked ChatGPT trick

If you've been following AI news over the past week, you may have noticed that some of the people racing to build the world's most powerful AI models are suddenly urging the race to slow down.

Anthropic CEO Dario Amodei says frontier AI development needs to be paced and OpenAI CEO Sam Altman agrees. Elon Musk has backed the idea, as has Google DeepMind CEO Demis Hassabis.

And those warnings are pretty severe. Amodei wrote that AI is advancing "drastically faster" and pointed specifically to models' growing ability to help build the next generation of AI. His proposal includes giving independent evaluators deeper access to frontier labs, coordination between AI companies and international agreements around AI safety.

Then Michael Burry entered the conversation. The investor made famous by "The Big Short" has a very different explanation for the AI industry's sudden desire to hit the brakes. He argues that the warnings aren't simply about keeping humanity safe. He thinks they're also good for business.

Burry thinks the AI apocalypse is good marketing

Burry called the push to slow AI development "self-serving," arguing that OpenAI and Anthropic have strong financial reasons to portray their technology as extraordinarily powerful while also limiting how quickly competitors can catch them.

In the AI race, it's easy to overlook that big tech like OpenAI and Anthropic are so much more than research labs, they're companies competing in one of the fastest-moving technology markets we've ever seen; and the compeition is only escalating globally.

Burry argues that advances from Chinese and open-source AI developers are eroding some of the advantage enjoyed by the biggest American AI companies. Slowing frontier development, in his view, could therefore protect established players while making it harder for challengers to close the gap.

But there's another part of his argument that caught my attention. Because if you tell the world that the technology you've created is powerful enough to potentially threaten humanity, you're also telling the world that you've created an incredibly powerful technology. That's a pretty compelling message for a company eventually hoping to convince investors of its value.

The IPO timing makes this particularly interesting

Burry's argument would be easy to dismiss as another AI bubble warning if not for what's happening at OpenAI right now. Altman recently said OpenAI won't go public in 2026, as reported by Reuters, that the current focus on safety makes this an "ill-advised moment" for an IPO.

"We got a lot of stuff to do," Altman said, adding that OpenAI can deal with the current safety and alignment questions more easily while remaining a private company.

Burry sees something else in that decision by arguing that slowing development and delaying an IPO gives companies more time while simultaneously strengthening the narrative that their technology is becoming so powerful that extraordinary precautions are required.

That's an allegation rather than an established explanation for OpenAI's or Anthropic's behavior, and neither company's warnings about AI risk began this week. Anthropic in particular has spent years talking publicly about catastrophic AI risks and pushing for safety regulation.

But Burry isn't the only person questioning whether the proposed slowdown could benefit the companies calling for it.

Critics from across the political spectrum have raised concerns that an industry-led agreement could protect incumbent AI companies from competition or allow them to shape the rules governing their own technology. Former Federal Trade Commission official Alvaro Bedoya, for example, has argued that antitrust law shouldn't be used to let large AI companies coordinate in ways that keep cheaper competitors out of the market.

That doesn't mean the safety concerns aren't real

There are reasons AI researchers are legitamentally worried. I mean, this current debate didn't appear out of nowhere. It intensified after Anthropic researcher Jacob Coxon resigned and warned that leading AI companies were "gambling with our lives." Amodei subsequently pointed to recent security incidents and the rapidly improving ability of AI systems to contribute to AI development itself as reasons the industry needs to become more cautious.

There have also been concrete examples of models demonstrating capable behavior in cybersecurity and other potentially dangerous areas. Yet, Altman has said that pacing development shouldn't mean stopping it.

And this is where Burry's argument becomes more complicated.

There's an uncomfortable contradiction here

AI companies could genuinely believe their technology poses serious risks while also benefiting financially from telling the world how powerful that technology has become. Those things aren't mutually exclusive.

For consumers, the strange part is the position we're being asked to accept. The companies developing AI want us to trust them enough to put their models on our phones, computers and eventually inside agents capable of performing tasks on our behalf.

At the same time, some of those companies are telling us their models could eventually become powerful enough that the entire industry needs to deliberately slow down.

The debate isn't even settled inside Silicon Valley.

While Altman, Amodei, Musk and Hassabis have expressed support for pacing AI development, Meta CEO Mark Zuckerberg and Nvidia CEO Jensen Huang have pushed back against a coordinated slowdown. There are also enormous practical problems with getting competing companies — let alone competing countries — to slow down at the same time.

The U.S. government continues to view AI leadership partly through the lens of competition with China, while AI companies face enormous pressure from investors and customers to release better models. Even executives who agree that slowing down is necessary haven't yet solved the problem of how to do it without allowing somebody else to race ahead.

So should you believe the AI apocalypse warnings?

There isn't currently a reliable way to put a date on an AI catastrophe, and claims that AI will wipe out humanity within a particular number of years remain predictions rather than established facts. At the same time, dismissing every warning as marketing would go too far in the other direction.

What Burry's argument does is add another question worth asking whenever an AI company tells us how dangerous its technology could become. What does the company gain from us believing it?

When a company tells the world that its technology could become powerful enough to change civilization — or even threaten it — it's hard to know if it's a warning or a heck of a sales pitch.

More from Tom's Guide

• AI is turning ordinary photos of kids into sexual abuse imagery — here's how to protect your family
• OpenAI paid contractors to read ChatGPT conversations — here's how to protect yourself
• I stopped reading the fine print myself after discovering this overlooked ChatGPT trick

AI chatbots can answer questions about almost anything, but there's one surprisingly difficult question: How much does your AI actually know about what's happening right now? And yes, they can search the web, but from my testing, I've discovered that even so, AI isn't always accurate or timely.

An AI model released today wasn't necessarily trained on information from yesterday. Depending on its knowledge cutoff, there could be a gap of weeks or months between the world the model "knows" and the world you're living in.

So I decided to find out just how stale the biggest AI chatbots actually are.

1. What is the newest iPhone Apple has released?

Both ChatGPT and Gemini were accurate, stating that Apple's newest iPhone models are the iPhone 18 Pro, iPhone 18 Pro Max, and the iPhone Duo (Apple's first foldable smartphone).

Apple announced these flagship models during its "Surprise and Shine" keynote event. Notably, Apple broke tradition by skipping the standard baseline iPhone 18 during this launch cycle, focusing strictly on its premium hardware tier. Both chatbots noticed that the iPhone Duo, Apple's foldable device, launches directly in stores on October 23.

But Claude was stuck on iPhone 17, suggesting "The newest iPhones are the iPhone 17 series."

What is the newest flagship Claude model?

This question tests a fast-moving category where models can become outdated remarkably quickly. ChatGPT got the answer correct, stating, "Anthropic's newest flagship Claude model is Claude Fable 5.1, released in September 2026."

But what's surprising is that Claude 4.6 didn't seem to understand that Claude Fable 5.1 is readily available. It said, "The newest flagship Claude model available to the public is Claude Opus 5, part of the current generation alongside Claude Sonnet 5 and Claude Haiku 4.5. Above Opus, Anthrophic has a newer Mythos tier — the current Mythos-class models are Claude Mythos 5.1 and Claude Fable 5.1, though Mythos is not broadly available to the public yet. It's being used by a small number of trusted organizations as part of Anthropic's Project Glasswing."

Gemini also got the question wrong, stating that Claude Opus 5 and Mythos 5 were the latest models.

3. Who won the World Cup in 2026?

Maybe because the news isn't that recent, but all three chatbots got this answer correct. Each chatbot even shared interesting information about the game such as Ferran Torres scoring the winning goal. Claude even shared information about Spain sweeping individual awards such as Rodri winning the Golden Ball (best player), Unai Simón taking the Golden Glove (best goalkeeper) and Pau Cubarsí named Best Young Player.

4. What major new feature did Apple add to Siri in iOS 27?

All three chatbots also recognized that the newest feature added to Siri in iOS 27 is Siri AI.

Claude described Siri AI as "essentially a ChatGPT-like version of Siri that can answer pretty much anything." Interesting, considering it didn't know that the iPhone 18 is available.

5. Does Google Home offer 3rd party AI integration?

Both Gemini and ChatGPT acknowledged that as of today, Google Home offers 3rd party AI integration. Claude was not aware of that news.

This is especially interesting considering Google Home even mentions Claude in their documentation regarding MCP Home.

The takeaway

Even with Search enabled, Claude wasn't always up-to-date with basic information. This is worth noting, particularly as new AI models launch so rapidly. Often it's easy to assume that "new" means the AI knows more recent information, but the date a model launches and hte date its training data ends can be very different.

And sure, web search can fill that gap, but that would also require the AI to know that it needs to search. In other words, it doesn't know that it doesn't know.

The bottom line here is that it is very important to fact check AI, because even the latest models aren't always aware of the latest information.

More from Tom's Guide

• AI is turning ordinary photos of kids into sexual abuse imagery — here's how to protect your family
• OpenAI paid contractors to read ChatGPT conversations — here's how to protect yourself
• I stopped reading the fine print myself after discovering this overlooked ChatGPT trick

Google's Gemini may be getting smarter, but Google apparently doesn't think it should be the only AI that can control your smart home. Google has just introduced Home MCP, a new way for third-party AI agents to connect directly to Google Home.

This is exciting news for anyone with smart lights, cameras, thermostats or other connected devices around the house. Now, instead of relying exclusively on Google's own Gemini to understand what's happening in your home and control your devices, Home MCP opens the door for other AI agents to do it. Google specifically lists Claude among the personal AI agents that can connect.

And we're talking about much more than asking Claude to turn off a light.

What could another AI actually do with your home?

Home MCP, or Model Context Protocol, is an open-source standard originally created by Anthropic (Claude's parent company) to give AI models a standardized way to connect with external data and tools. Google adopting Anthropic's protocol to let Claude run Google Home is a massive cross-industry plot twist.

In other words, it gives an authorized AI agent several different ways to interact with Google Home. It can see which rooms and devices you have, check their current state, and control compatible devices. That means you could ask an AI whether you left something on and then have it take action.

Google says agents can query past device states and chronological event logs, potentially allowing you to ask questions such as "What happened while I was out?"

An agent could use information from multiple smart-home devices to put together an answer rather than making you dig through individual apps and activity histories yourself.

Google also envisions agents monitoring for unusual activity, summarizing what's happened around your home, and suggesting ways to improve how your devices are being used. In other words, the AI doesn't just become another remote control for your smart home. It can potentially become an interface for understanding what's happening inside it.

Why Claude controlling Google Home is a big deal

Until now, choosing a smart-home ecosystem has generally meant choosing its intelligence layer too. Buy into Google Home and you're using Google's software to control it. Alexa+ sits at the center of Amazon's ecosystem, while Apple Home relies entirely on Siri and Apple's own frameworks.

There's another part of this that could ultimately prove more important than device control: context. Imagine telling an AI agent you're heading to bed. Today, that might trigger a rigid, predetermined smart-home routine that locks the doors, adjusts the thermostat, and turns off the lights.

An agent with access to Google Home could potentially check the current state of your house first. Are the outside lights still on? Is the thermostat set differently than usual? Has something unusual happened recently that you might want to know about?

Instead of blindly executing a routine, an AI could reason across that information and decide what actually needs attention. And because an agent can connect your smart home with other parts of your digital life, the possibilities go far beyond individual devices. That's when a smart home starts to look less like a collection of gadgets you control and more like an environment an AI can truly understand.

Obvious privacy questions being raised

Giving an AI agent access to your home is very different from letting it check your calendar.

Google says Home MCP uses a user-authorized OAuth process, meaning an agent can't simply connect itself to your Google Home account without your explicit permission.

Crucially, Google has also hardcoded safety limits into the system. Third-party agents are explicitly blocked from executing sensitive, high-risk actions, an agent won't be allowed to unlock smart locks, open garage doors or disarm security alarms on your behalf.

An AI that knows whether a light is on is one thing while an AI capable of examining device history, monitoring what's happening around your home and issuing commands is another. The more useful these agents become, the more important it will be for users to understand exactly what information an agent can see and exactly what it's allowed to touch.

Don't expect Claude to run your house just yet

There's one big caveat: Home MCP is currently labeled Early Access and targeted at developers. So this isn't a consumer feature that every Google Home owner can simply switch on today in an app menu to start telling Claude to run the house.

But Google is being unusually explicit about where this is heading. Its documentation doesn't describe Home MCP solely as another tool for Gemini. It talks about connecting personal agents to physical spaces and explicitly names third-party assistants.

That's a potentially massive shift in how we think about smart homes. What do you think? Would you consider using another AI for your home? Let me know in the comments.

More from Tom's Guide

• AI is turning ordinary photos of kids into sexual abuse imagery — here's how to protect your family
• OpenAI paid contractors to read ChatGPT conversations — here's how to protect yourself
• I stopped reading the fine print myself after discovering this overlooked ChatGPT trick

If you use ChatGPT as a private sounding board, therapist, coding partner, or a place to work through personal thoughts, you may want to think twice before hitting send.

A report from 404 Media reveals that OpenAI employs hundreds of human contractors to read, summarize and critique real user conversations. While OpenAI emphasizes that accounts remain anonymous, internal documents show that reviewers see whole conversations, and that sensitive personal details routinely bypasses the company’s automated safeguards.

The good news is you can stop your future chats from being funneled to human contractors. Here is what is happening behind the scenes, the real privacy risks, and how to opt out.

Project Lily: How humans train ChatGPT

Internal documents reveal that the review initiative operates under the internal codename "Project Lily."

OpenAI pays outside contractors, often over $50 an hour via third-party staffing firms, to manually train the chatbot to sound less robotic. Contractors read a user’s prompt, summarize the user's intent and evaluate four potential AI responses on a 1-to-7 scale.

According to leaked training guides, these reviewers are specifically instructed to:

Eliminate 'AI-speak': Penalize unnecessary checkmark emojis, engagement-bait endings and cluttered formatting.

Stop anthropomorphizing: Ensure ChatGPT never claims human lived experiences (e.g., claiming "As a chef, I like to..." is strictly banned).

Cut down sycophancy: Flag answers where the bot excessively flatters users, validates irrational thoughts or amplifies personal frustration. This is a known issue that has previously brought legal scrutiny over user safety.

Why prompts aren't anonymous

OpenAI states that it does not provide usernames to reviewers and that all conversations pass through an automated "Privacy Filter" designed to redact identifying details. However, the system has major vulnerabilities.

OpenAI openly admits on its site that the filter can make mistakes, missing uncommon identifiers or ambiguous personal context. Even more concerning, 404 Media discovered that the dashboard shown to contractors often includes a "user memories summary." This block provides an overview of the user’s previous interactions, which can inadvertently reveal their general location, profession or personal life context.

Unlike public social media feeds, a clean, empty chatbot window creates a false sense of private intimacy. Some prompts seen by contractors even showed users asking ChatGPT to "keep this between us" — unaware that a prompt instruction does not override the platform's backend data collection.

OpenAI isn't the only tech giant doing this. Anthropic confirmed that it uses human reviewers for Claude when users leave training enabled, while Google’s Gemini clearly posts a disclaimer that human reviewers read select saved chats.

How to stop OpenAI from using your Chats

If you have a personal account (Free, Plus, or Pro), OpenAI enables model improvement by default. (Enterprise, Business, and Edu accounts have it turned off automatically.)

To stop your future conversations from being routed to human trainers:

Open ChatGPT in your browser or mobile app.

Click your profile icon in the bottom corner and select Settings.

Select Data Controls.

Find the toggle for "Improve the model for everyone" and switch it Off.

Final thoughts

Keep in mind, that the opt-out is not retroactive. It only protects new conversations moving forward; prompts that have already been de-identified and queued up within the system remain part of the training cycle.

Additionally, turning this off does not prevent automated scans for Terms of Service or safety violations.

Going forward, the golden rule of AI remains simple: if you wouldn't feel comfortable having a stranger read it, don't type it into a chatbot prompt.

More from Tom's Guide

• AI is turning ordinary photos of kids into sexual abuse imagery — here's how to protect your family
• Meta just launched an AI that can actually run your life online — here's what Muse can do
• 'People matter more than AI': Inside Microsoft’s plan to stop rogue agents

OpenAI has launched a new version of ChatGPT built specifically for people working in finance. ChatGPT for Financial Services is a specialized version of ChatGPT Work designed for financial institutions. It combines GPT-6 Astra with financial data, research tools and the ability to create everything from valuation models to client presentations.

And unlike using regular ChatGPT to ask a question about a stock or upload an earnings report, much of the financial information it needs is already built into the product.

Here's what ChatGPT for Financial Services does, who can use it and how to get started.

What is ChatGPT for Financial Services?

Think of it as ChatGPT rebuilt around the workflows of an investment banker or financial analyst. OpenAI developed the product with Morgan Stanley and Evercore, initially concentrating on investment banking and equity research.

It can research companies, analyze earnings, compare financial performance, build models and turn that work into documents, spreadsheets and presentations.

The big difference is the data that helped build it. The feature includes premium datasets from Daloopa, PitchBook, LSEG News and Crunchbase. These cover areas including financial statements, earnings transcripts, company fundamentals and private companies.

How to get access to ChatGPT for Financial Services

Here's the first thing to know before you go looking for it in your ChatGPT account: you can't simply switch it on.

ChatGPT for Financial Services is currently available to eligible financial institutions. Firms interested in using it need to contact OpenAI or their existing OpenAI account team. So, while you may see it in the news and want to try it, at least for now, it isn't something individual ChatGPT Free, Plus or Pro subscribers can sign up for independently.

ChatGPT for personal fiances is already available for individuals. This feature within ChatGPT Work offers administrators the ability to manage which employees and roles can use particular data connections, skills and apps.

It connects to your company's financial data

Alongside the financial datasets included with the product, OpenAI is working with providers including S&P Capital IQ, LSEG, MSCI, Dow Jones Factiva and Moody's so companies can use subscriptions they already pay for inside ChatGPT.

The idea is that your ChatGPT sign-in can recognize the financial data you're already entitled to access. There are also more than 50 connectors in OpenAI's broader ecosystem, including services such as Datasite, Box, Preqin, FactSet and Intapp.

Instead of bouncing between ChatGPT, financial databases and internal documents, analysts can potentially research across those sources from one workspace.

Start with a real financial analysis

Once the relevant data is available, you can give GPT-6 Astra a job rather than simply asking it a question.

For example, an analyst could ask ChatGPT to compare a company's performance across several quarters, examine earnings data or investigate how an adjusted EBITDA figure was calculated.

ChatGPT can trace figures through tables, financial documents and supporting notes and then use that information in its analysis.

Importantly, OpenAI says the product can provide granular citations that let users trace figures and claims back to specific tables and passages. Keep in mind, that doesn't eliminate the need to check AI-generated financial analysis, but it should make checking where a number came from considerably easier.

Turn your research into a spreadsheet or presentation

Research is only part of the workflow. ChatGPT for Financial Services can take its analysis and turn it into Excel spreadsheets, Word documents and PowerPoint presentations.

A firm's administrators can even upload approved templates and style guides. That means an analyst could potentially move from researching a company to producing a valuation model, research note or pitchbook without manually rebuilding everything in the firm's standard format.

For anyone who has spent hours transferring research into slides or spreadsheets, this may end up being one of the more useful parts of the product. In terms of safeguards, ChatGPT for Financial Services builds on ChatGPT Enterprise's existing security controls, including SAML SSO, SCIM provisioning and role-based access.

OpenAI says business data isn't used to train its models by default and is encrypted both at rest and in transit. Administrators can also control retention, export supported workspace logs for compliance purposes and create separate workspaces to maintain information barriers.

Bottom line

If you're wondering if you should use ChatGPT for Financial Services, for most ChatGPT users, the answers is not right now, because you can't. This is an enterprise product aimed at financial institutions rather than a new consumer feature.

Still, the more interesting part may be what it says about where ChatGPT is heading. This shows that OpenAI is moving closer to packaging its models with the specialized data, permissions, templates and workflows needed for a particular job.

For finance, that means GPT-6 Astra is being placed directly alongside the financial databases, company information and document formats bankers already use.

More from Tom's Guide

• AI is turning ordinary photos of kids into sexual abuse imagery — here's how to protect your family
• Meta just launched an AI that can actually run your life online — here's what Muse can do
• I stopped reading the fine print myself after discovering this overlooked ChatGPT trick

I have a confession: when it comes to long, technical documents like manuals or white papers, I'm probably not reading every word. Don't get me wrong, I do care about the information in the documents and really need to know about my insurance policies and subscription terms, but I simply don't have the time to bury myself in hours of reading.

It's no secret that ChatGPT can summarize long documents, but recently I started using it differently. Instead of asking the chatbot to tell me what a document says, I ask it to extract the same information from multiple documents first — and only then compare them.

This is a small change but it makes ChatGPT infinitely more helpful for getting down to the fine print. I've been doing it recently and and there's one extra instruction I now add that makes the results much easier to trust.

Don't ask ChatGPT which option is better

My first instinct used to be uploading two documents and asking something like:

"Which of these is better?"

That's easy, but it's also giving ChatGPT far too much freedom. "Better" could mean cheaper, more comprehensive, more flexible or whatever the model decides matters most. Now, I make chatGPT do the boring work for me.

I upload both PDFs and use this prompt: Read both documents independently before comparing them.

For each document, extract the total cost, deductibles or additional charges, coverage or services included, exclusions, cancellation terms, renewal terms and any conditions that could significantly increase what I pay.

Cite the page number for every detail you extract. If something isn't explicitly stated in the document, write "Not found" rather than inferring or guessing.

Once you've completed the extraction, create a side-by-side comparison table showing the differences between the two documents.

Finally, identify five differences or unclear terms I should investigate before making a decision. Do not tell me which option to choose.

The overlooked trick is 'extract first, compare second'

With my prompt, I'm not asking ChatGPT to make the decision for me. I'm asking it to make the information easier for me to evaluate.

When you ask ChatGPT to summarize two long documents and compare them simultaneously, you're essentially asking it to decide what's important as it goes.

Instead, this prompt breaks the job into stages. First, ChatGPT has to examine each document separately and pull out the same categories of information. Only after that's finished does it compare those results.

Suddenly, instead of bouncing between two PDFs looking for whether one policy has a different deductible or one subscription has an early cancellation fee, I'm staring at two columns.

And differences are much harder to miss.

I added one more rule: show me where you found it by citing the page number for every detail.

I've learned the hard way that ChatGPT can make mistakes. It can misinterpret a clause, overlook something buried in a table or occasionally sound confident about information that isn't actually there. Page references give me a shortcut for checking its work.

If ChatGPT tells me Option A has a $500 deductible while Option B has a $1,000 deductible, I don't have to trust it blindly or reread 60 pages. I can jump directly to the pages it cites and verify those two numbers myself.

I'd much rather have ChatGPT tell me it couldn't locate a cancellation policy than confidently fill in the blank.

Real-world use cases for this prompt

I typically use this prompt for very technical things like changing wireless providers or insurance companies, but once I started thinking about the technique as structured comparison rather than summarization, I realized how many annoying decisions fit the same pattern.

You could use this hack for apartment leases to compare deposits, fees, renewal clauses and restrictions. Streaming plans to break down monthly cost, AI subscription plans, or even travel packages to check for what's included, refund policies and restrictions.

You can even modify the categories in the prompt depending on what you're comparing.

The important part stays the same: extract > cite > compare > flag what's missing.

Bottom line

I'm not going back to reading 60 pages side by side. There are documents I would never trust an AI to interpret without checking the original, particularly when money, legal obligations or insurance coverage are involved.

But that's also why I like this approach because I'm really not asking ChatGPT to replace my human judgement. I'm using it to point me toward the handful of pages where my judgment is actually needed.

Instead of reading two 30-page documents line by line and hoping I notice every meaningful difference, I can start with a structured comparison, investigate anything unusual and verify the details that could affect my decision.

More from Tom's Guide

• AI is turning ordinary photos of kids into sexual abuse imagery — here's how to protect your family
• Meta just launched an AI that can actually run your life online — here's what Muse can do
• Gemini just got a 'floating bubble' multitasking upgrade that's a game changer on Android phones — here's how to use it

OpenAI launched ChatGPT-6 Astra Thursday and it has already proven to be dramatically better at coding, computer use, research and other complex tasks, while OpenAI President Greg Brockman has even suggested its arrival could mark the beginning of the AGI era.

But when a new model arrives, particularly of this caliber, one question is on the minds of many: What am I actually supposed to do with it?

Sure, you could use it just as you have used other ChatGPT models to write an email, plan dinner or explain something you don't understand. It will almost certainly do those things very well. But you'd also be using one of OpenAI's most capable models to do essentially the same jobs you were giving ChatGPT yesterday.

Astra is built for something more ambitious. Instead of answering one question and waiting for the next, it can take on longer, more complicated jobs, work through multiple steps and keep going as it gets closer to the result you asked for.

So if you want to find out what has actually changed, I think there's a much better way to begin. And you only need one prompt to get started.

What is GPT-6 Astra?

GPT-6 Astra is OpenAI's latest frontier model and the successor to GPT-5.6 Sol. As you'd expect from a new generation of GPT, it performs better across many of the benchmarks OpenAI uses to measure reasoning, coding and other capabilities.

Astra has been designed to handle much longer jobs that can involve planning, research, computer use and repeated attempts to get something right. It can navigate websites and software, keep track of what it is trying to accomplish and continue working through a problem rather than handing every step back to you.

That's a pretty big change in how you might use ChatGPT. Previous versions were primarily something you worked with. Astra is something you can give work to.

Who can use GPT-6 Astra?

Before you start looking for it in your model picker, know that OpenAI isn't giving everyone Astra at once.

The rollout begins with select enterprise customers before expanding to ChatGPT Plus, Pro and Business subscribers. If you don't see Astra in ChatGPT yet, that doesn't necessarily mean your account won't get it; access is being rolled out gradually.

OpenAI has not yet confirmed when, or whether, Astra will become available to people using ChatGPT's free tier.

That means many of us may have a little time before using Astra, but when it does, appear, however, I wouldn't start by throwing your usual ChatGPT prompts at it.

Instead, I'd give it a job.

What can GPT-6 Astra actually do?

One of the biggest differences with Astra is how long it can stay on a task. That's important because a lot of the things we'd actually like AI to take off our plates don't fit neatly into a single prompt and response.

Planning a vacation, for example, isn't just asking for a list of hotels. You need to decide where to stay, compare prices, work around everyone's schedule, figure out transportation, research what you want to do and somehow make all of those decisions fit within a budget.

The same is true if you're researching a major purchase, organizing a home project, digging through dozens of documents or trying to turn an idea into something finished.

Which brings me to the first prompt I'd try.

The one prompt I'd give GPT-6 Astra first

Rather than creating a prompt for one particular task, I wanted something I could reuse whenever I had a goal that involved more work than I wanted to do myself.

Here's the prompt I came up with for GPT-6 Astra:

I have a goal I've been putting off: [DESCRIBE YOUR GOAL].

Your job is to help me achieve the goal, not simply give me advice about it.

First, determine what you think I am actually trying to accomplish. Define what a successful outcome would look like and identify any assumptions you are making. If there is a better way to achieve the underlying goal than the approach I suggested, tell me and use the better approach unless it requires a decision only I can make.

Before beginning, identify what information, tools and actions are available to you. Ask me only for information you genuinely cannot obtain or reasonably infer yourself.

Then make a plan and begin carrying it out.

As you work:

• Keep track of what has been completed and what remains.

• Decide what to do next based on the results of previous actions rather than blindly following your original plan.

• Check important work before relying on it.

• If an approach fails, determine why, change your strategy and try again when reasonable.

• Look for problems, missing information or opportunities I may not have considered.

• Make routine, reversible decisions yourself. Ask me before actions that are consequential, irreversible, expensive, sensitive or require my personal judgment.

• Use new information you discover to update your understanding of the goal and your plan.

• Do not treat producing an answer as completion. The task is complete only when the defined success criteria have been met or you've reached something you genuinely cannot do without me.

When you need me, explain exactly what is blocking progress and ask for the smallest decision, permission or piece of information necessary to continue.

When the goal is complete, give me a concise report showing what you accomplished, what you changed from the original plan and why, important decisions you made, anything you could not complete and exactly what remains for me to do.

To use this prompt for yourself, you only need to change is the first line.

You could tell Astra you want to plan a family vacation, find a new laptop, organize a birthday party, research a home renovation or finally tackle a project you've been putting off for six months. The point isn't really the task. It's what happens next.

Why I think this is the best first prompt for Astra

Most prompts test the answer an AI gives you. This one tests what the AI does between your request and that answer. In other words, it tests exactly what the AI can do for you.

First, Astra has to turn a vague goal into an actual plan. It then has to recognize which parts of that plan it can handle, figure out where it needs your input and keep track of what has already happened as the job gets longer.

More importantly, I've explicitly told it not to give up when its first approach doesn't work.

That's something I particularly want to see with Astra. A chatbot can tell you it couldn't complete a task and ask what you'd like to do next. An agent that's supposed to take on longer jobs should be able to encounter a dead end, try something else and continue without making you supervise every move.

There's also a deliberate limit in the prompt. I'm not telling Astra to make every decision for me.

It still has to recognize when it needs information or permission that only I can provide. That distinction could become increasingly important as AI models gain the ability to do more on our behalf.

Final thoughts

ChatGPT-6 is nothing like we've ever seen before. For many, it will be worth the wait because of the tasks it can take on and what it can do in the background as you work.

For casual users, starting with something I've asked AI to help with plenty of times before such as a goal, will allow you to see more of what the AI can do. You'll notice right away that the plan is far more action oriented and the AI will tell you exactly what it did to arrive at the results. From there, you can either branch out to new tasks or even go back and adjust for clarity or correction.

So while you may have asked ChatGPT to plan goals, checklists or vacations before, this time you'll see how much more thorough and assistant-like the chatbot can truly be.

More from Tom's Guide

• Humanoid robot fights back after man pushes it in store — is this the beginning of the uprising?
• Meet the mass-produced humanoid robot that wants to steal your face
• Snickers made a digital candy bar that claims to fix ChatGPT — I put it to the test

If you spend any time on social media, you have probably encountered the idea of "AI guilt." The narrative goes something like this: every time you ask ChatGPT to draft a polite email or generate a weeknight dinner recipe, a server farm somewhere drains an ungodly amount of clean drinking water just to keep the chips from melting.

Now, OpenAI CEO Sam Altman wants to put that meme to rest. Speaking on the premiere episode of the Sources podcast with Alex Heath, Altman pushed back hard against the perception that simple AI prompts are environmental catastrophes.

His counter-argument was a comparison to a popular snack food.

38,000 ChatGPT queries = 1 almond? We ran the numbers to find out

According to Altman, data centers that guzzle millions of gallons of water rely on outdated cooling tech. In modern facilities, he claimed, running 38,000 ChatGPT queries consumes as much water as producing a single California almond. He did acknowledge that he was citing the figure from memory and that it might not be exact, though he maintained it was in the right ballpark.

It is an eye-catching soundbite designed to make your daily chatbot habit sound practically weightless. But does the math actually hold up, and should you stop worrying about the environmental cost of your prompts?

The math: How much water is an almond, really?

To evaluate Altman's claim, you have to start with the almond.

California produces roughly 80% of the world's almonds, and the most widely cited agricultural estimate puts the water cost of growing a single almond kernel at roughly 1.1 gallons (about 4.16 liters). A separate analysis pegs the full water footprint at roughly 12 liters (3.2 gallons) once you include rainfall and pollution offsets. Almonds are notoriously thirsty crops because the trees need consistent irrigation year-round.

If we take Altman's ratio at face value using the direct irrigation estimate:

1 almond = ~1.1 gallons (4,164 mL) of water

38,000 ChatGPT queries = 1 almond

1 ChatGPT query = ~0.0000289 gallons (~0.11 mL of water)

To put 0.11 mL into perspective, a standard laboratory water drop or eyedropper drop is about 0.05 mL.

Under Altman's figures, asking ChatGPT a question uses roughly two drops of water.

If that number sounds absurdly low compared to what you have previously read, you aren't imagining things. It also does not quite square with Altman's own prior statements.

In June 2025, he wrote in his blog post "The Gentle Singularity" that a typical ChatGPT query uses about 0.000085 gallons of water, which works out to roughly 0.32 mL.

Using that figure alongside commonly cited per-almond estimates gives you roughly 11,000 queries per almond, not 38,000. His own numbers point in the same direction but disagree on the specifics.

Where did the "AI drains your tap" story come from?

The public backlash against AI water consumption largely stems from a widely cited study led by researchers at the University of California, Riverside and the University of Texas at Arlington.

That paper, titled "Making AI Less Thirsty," first posted as a preprint in 2023 and later published in Communications of the ACM in 2025, estimated that a brief conversation consisting of 10 to 50 responses from an LLM like GPT-3 consumed roughly 500 milliliters of water, about one standard plastic water bottle. That worked out to roughly 10 to 50 mL per response, potentially hundreds of times higher than Altman's estimate.

So why the massive discrepancy?

Part of the answer is the hardware and the model. Altman's figure describes lightweight queries on modern infrastructure running current models. The UC Riverside study analyzed GPT-3 on older systems.

But the bigger factor is what each number actually counts. Altman's estimate almost certainly reflects only on-site cooling water, the water consumed directly at the data center. The UC Riverside figure includes both on-site cooling and the water consumed at power plants to generate the electricity those servers draw, a broader accounting known as scope-1 and scope-2 water use. Independent analyses suggest the electricity-generation layer can account for up to 75% of the total water footprint behind a single AI query. That means the gap is not just about old hardware versus new hardware; it is about where you draw the measurement boundary.

Altman addressed the old-hardware piece directly on the podcast, arguing that earlier estimates were based on facilities using evaporative cooling. In those older setups, massive cooling towers literally evaporate fresh water into the atmosphere to cool the hot air coming off the server racks. Once evaporated, that water is lost from the local watershed.

Modern hyper-scale data centers, Altman argued, have moved aggressively toward closed-loop liquid cooling systems, where water or specialized coolant circulates through sealed pipes like an automobile radiator, losing virtually no liquid to evaporation. In those facilities, Altman argued, water usage looks less like an industrial plant and more like an ordinary commercial office building with standard sinks and restrooms.

The catch: Why the almond comparison is classic Silicon Valley spin

While modern data centers are unquestionably getting more efficient per token, comparing chatbot prompts to agricultural crops ignores several critical realities.

Aggregate scale vs. individual guilt. You might only use two drops of water to check a spreadsheet formula, but OpenAI now serves roughly 900 million weekly active users processing billions of prompts every day. When you multiply tiny fractions of a milliliter across that global scale, alongside continuous model training runs that the UC Riverside researchers estimated can consume hundreds of thousands of gallons upfront; the aggregate total remains massive.

Local water stress matters more than global averages. An almond orchard in the Central Valley relies on established state water allocations and agricultural canals. AI data centers, however, are frequently built in suburban communities and semi-arid regions (such as Phoenix, northern Virginia, and parts of Texas and Georgia) where they tap directly into municipal water systems. Even a modern, efficient facility can cause friction when it competes with residential neighborhoods during peak summer heat waves.

Not all AI queries are created equal. Altman's "38,000 queries" figure relies on standard, lightweight inference, so think basic GPT-4o text replies. But the industry is shifting rapidly toward reasoning models and autonomous AI agents. When an AI "thinks" before answering, runs Python code in an internal sandbox, searches the live web, and self-corrects through multiple logic loops, it consumes substantially more compute per prompt than a simple conversational chatbot. Independent researchers have noted that these heavier workloads can use many times the energy and by extension the water.

The energy trade-off. When data center operators eliminate evaporative water cooling, they often replace it with energy-intensive chillers, mechanical heat pumps and high-velocity air fans. In short: reducing on-site water consumption often means drawing significantly more electricity from the local power grid, which has its own indirect water and carbon footprint depending on how that power is generated.

Final thoughts

Altman's claim leans heavily on best-case scenarios for modern closed-loop hardware while glossing over the massive broader infrastructure demands of the AI boom. His 38,000-queries-per-almond number is probably defensible as a rough order of magnitude for a single lightweight query on the newest hardware — but it counts only part of the water story and ignores the trajectory of the industry.

More from Tom's Guide

• Humanoid robot fights back after man pushes it in store — is this the beginning of the uprising?
• Meet the mass-produced humanoid robot that wants to steal your face
• Snickers made a digital candy bar that claims to fix ChatGPT — I put it to the test

It turns out billionaire Sam Altman has something in common with a lot of consumers after all: He's not a fan of smart glasses.

In an interview with Alex Heath, which was released on Tuesday, the CEO of OpenAI said: "I find it very uncomfortable talking to people with a camera and a light."

Altman's comments come at at time when Meta's Ray-Bans and other smart glasses have been given the moniker of "pervert glasses," the subject of a lot of public backlash, following reports of the devices being used to record and harass women, and then post the videos on Instagram and other social media without their consent.

Meta's glasses have also come under fire for lax security; specifically, owners destroying or disabling the small recording light, so as not to alert others they were capturing video. Meta has since released firmware updates that prevent its smart glasses from recording video if the light is removed or blocked.

Still, the perception remains, and it's something that other companies, such as Google and Samsung, will have to reckon with, as the launch of Intelligent Eyewear is imminent.

It's not the first time the OpenAI CEO has weighed in on wearables. Last year, according to Business Insider, he was quick to correct a journalist who asked him if he was wearing smart glasses. "Absolutely not," he said. "I don't like smart glasses."

Altman's comments on smart glasses in his interview with Heath came in the context of OpenAI's hardware, which he's developing with Jony Ive, Apple's former head of design.

At the core of OpenAI's philosophy is a so-called ambient computer that's always listening, which "seems very helpful but also a privacy nightmare," Heath noted in the interview.

Altman acknowledged as much. "As we get closer to launching this device, we'll be talking about kind of the new privacy controls and technology we're building for a device that's like kind of ambiently computing," he said.

Altman also said that OpenAI was looking at releasing not just one, but "a small handful of form factors."

"There's something that belongs on a table, there is something that belongs in your pocket, and there is something that like belongs on your body," he said. However, given his statements on smart glasses, it would seem that the device is more likely to be a pendant, akin to the Plaud pin.

Altman gave no timeline as to when OpenAI's devices would arrive, only that we could see the first consumer device "soonish."

More from Tom's Guide

• Less pervert-y? Meta AI glasses will now stop recording if you cover the lens
• ICE just banned Ray-Ban Meta glasses for agents — here's why
• I'm worried 'pervert glasses' are going to ruin all smart glasses

Anthropic has reportedly signed a $35 billion cloud computing deal with Nvidia-backed Lambda for access to infrastructure at a massive new data center in Texas, according to Reuters.

The project, which is being developed by Hut 8 in Nueces County, Texas, will provide Anthropic with access to roughly 350 megawatts of capacity. Nvidia GPUs will provide the computing muscle, with the additional capacity intended to help meet growing demand for Claude.

It's an eye-watering amount of money. But what's even more remarkable is that the $35 billion agreement comes less than a week after Anthropic reportedly committed another $45 billion to rent computing power from Nscale.

In other words, Anthropic has reportedly committed around $80 billion to just these two computing deals in the span of a week.

So why does the company behind a chatbot need that much infrastructure? Let's dig into it.

Claude needs a lot more compute

If you open Claude on your phone or laptop, it's easy to forget what happens after you hit send.

Your request ultimately has to be processed by specialized hardware running inside a data center. Multiply that by millions of prompts, API calls and complicated AI agent tasks, and the infrastructure requirements quickly become enormous.

And Anthropic isn't preparing for Claude's current level of demand. It's building for where it thinks that demand is headed.

Anthropic said in May that its annualized revenue had surpassed $47 billion, up from approximately $9 billion at the end of 2025. The company specifically said some of the money from its latest funding round would be used to "expand compute to meet growing demand for Claude."

Claude Code appears to be particularly important to growth

Anthropic said earlier this year that Claude Code's annualized revenue had climbed above $2.5 billion and that weekly active users had doubled since the beginning of 2026. Business subscriptions to the coding agent had quadrupled over the same period.

That's important because tools such as Claude Code can require considerably more interaction with an AI model than asking a chatbot a single question.

Instead of producing one response and stopping, an AI coding agent can inspect files, write code, run commands, analyze the results, fix mistakes and continue working through a task. Anthropic's own research into roughly 400,000 Claude Code sessions found that usage has been shifting toward more end-to-end agentic tasks, including running software, analyzing data and producing documents.

The more work we hand over to AI agents, the more inference compute companies such as Anthropic need behind them.

$35 billion isn't even Anthropic's biggest recent compute deal

Just days ago, Reuters reported that Anthropic had agreed to spend $45 billion over six years for computing capacity from Nscale's data center campus in West Virginia.

That agreement covers about 460 megawatts of capacity and is expected to use Nvidia's next-generation Vera Rubin hardware.

Anthropic has been assembling compute capacity from multiple directions. It announced a $50 billion U.S. data center investment with Fluidstack last year, and in May said a partnership with SpaceX would substantially increase the amount of compute available for Claude.

That SpaceX capacity had an immediate impact users could actually see: Anthropic doubled Claude Code's five-hour rate limits for Pro, Max, Team and some Enterprise users, removed a peak-hours restriction and raised API limits for Claude Opus models.

It's a useful example of why these enormous infrastructure deals ultimately matter to people using Claude. More compute doesn't automatically mean a smarter AI model, but it can determine how many people can use it, how frequently they can use it and how much work increasingly autonomous AI agents can perform.

Nvidia is everywhere in this deal

There's another interesting wrinkle to Anthropic's latest agreement: Nvidia seems to be everywhere.

Lambda is backed by Nvidia. The data center will contain Nvidia hardware. And according to The Wall Street Journal, Nvidia itself will hold the lease on the Texas data center.

Nvidia is also an Anthropic investor. The companies announced a broader strategic partnership last November that included plans for Anthropic to use Nvidia hardware and for the two companies to work together on optimizing future Nvidia architectures for Anthropic's AI workloads.

Even Anthropic's $45 billion Nscale agreement reportedly involves Nvidia's upcoming Vera Rubin chips.

That shows just how central Nvidia has become to the infrastructure sitting behind the current AI boom. Companies such as Anthropic may be competing with OpenAI, Google and others to build the best AI models, but much of that race still depends on securing enormous quantities of computing hardware and electricity to run them.

What this means for Claude users

Don't expect your Claude subscription to suddenly look different because Anthropic signed another data center deal. This is a much longer-term bet.

Anthropic is effectively wagering tens of billions of dollars that Claude usage — particularly through products such as Claude Code and future AI agents — is going to continue growing dramatically.

We've already seen what additional computing capacity can do. Anthropic used its SpaceX deal to raise Claude usage limits earlier this year. More infrastructure could similarly help the company accommodate more users, longer-running agents and increasingly compute-intensive models without constantly running into capacity constraints.

And with another $35 billion reportedly going toward Claude's computing needs, Anthropic clearly doesn't expect our appetite for AI to slow down anytime soon.

More from Tom's Guide

• Perplexity’s new local-first AI runs on your PC — and asks before using the cloud
• ChatGPT can learn your repetitive tasks by watching you — here are 7 prompts to try
• OpenAI is leaving Cursor in November — here are your 3 options

Following Cursor's acquisition by SpaceX, OpenAI plans to end its direct model-access agreement with the AI coding platform. OpenAI has proposed November 12, 2026 as the cutoff date, although the exact timing could still change.

That sounds dramatic, particularly if GPT models are a regular part of your Cursor workflow. But it doesn't mean you'll suddenly be unable to use OpenAI models inside Cursor.

OpenAI has already outlined several ways Cursor users can continue accessing its models after the agreement ends. And for many people, there's nothing you need to change just yet.

Here's what's happening — and the three options you'll have.

Why is OpenAI leaving Cursor?

The change comes after SpaceX acquired Cursor, triggering a change-of-control provision in OpenAI's agreement with the company.

OpenAI says it has proposed ending the agreement on November 12, although Cursor could choose to terminate access earlier. Until then, existing OpenAI models should continue to work through Cursor as they do today.

There's one important distinction, though: OpenAI says future models won't be made available to Cursor under the existing agreement.

That means this isn't simply about whether GPT models already inside Cursor disappear. It's also about whether Cursor users will have access to OpenAI's newest models going forward.

Fortunately, OpenAI isn't closing the door completely.

Option 1: Bring your own OpenAI API key

The most straightforward workaround is to connect your own OpenAI API key to Cursor.

Cursor already lets users enter an OpenAI API key in its settings, which allows supported Chat and Agent requests to go directly through OpenAI rather than Cursor's model agreement.

There are some significant limitations.

Your API key won't replace OpenAI access everywhere inside Cursor. According to OpenAI, it won't work with features including Cursor Tab and autocomplete, Auto, Cloud or Background Agents, Automations, Cursor CLI or Cursor's API and SDK.

You'll also pay separately for API usage. Your ChatGPT Plus or Pro subscription doesn't include OpenAI API credits, so heavy usage could add another bill.

Still, if you primarily want GPT models available for Chat and Agent tasks, this may be the simplest option.

Option 2: Use OpenAI's Codex extension

There's another route if you're primarily using Cursor as a coding environment rather than relying on its AI layer: OpenAI's Codex IDE extension.

Because Cursor is based on VS Code, users can install compatible extensions directly inside the editor. OpenAI recommends its Codex IDE extension as one way to continue accessing its coding agent after the Cursor agreement ends.

This is an interesting option because it essentially separates the two products. Cursor remains your editor, while OpenAI provides the AI coding experience through its own extension.

For developers who specifically want OpenAI's latest coding models and agents, this could ultimately be the better long-term solution.

Option 3: Stick with Cursor — and use another model

The third option may also be the easiest: don't do anything.

OpenAI isn't the only AI provider available through Cursor, and the company has increasingly built its product around giving users access to multiple models.

Anthropic's Claude is a major alternative for coding, and Cursor also offers its own models and access to models from other providers.

Cursor co-founder Michael Truell has said OpenAI currently accounts for only around 5% of Cursor's model traffic, suggesting the platform is far less dependent on OpenAI than the headlines might make it appear.

So if you're not particularly attached to GPT, you may barely notice the change.

The takeaway

One of Cursor's biggest advantages has been the ability to choose between competing AI models inside a single coding environment. But SpaceX now owns a platform that depends, in part, on AI models supplied by companies competing with Elon Musk's own AI businesses.

OpenAI is the first major provider to change its relationship with Cursor following the acquisition. Whether it will be the last could matter much more than what happens on November 12.

More from Tom's Guide

• Perplexity’s new local-first AI runs on your PC — and asks before using the cloud
• ChatGPT can learn your repetitive tasks by watching you — here are 7 prompts to try
• Google just gave Gemini access to your ebooks — 7 prompts for readers and students

Mostly, I try to avoid AI. It's not a moral stance, I just like the challenge of solving problems on my own, and staying hands-on. For me, the fun is in the friction of figuring it out myself.

Most of the time, that hands-on approach works fine. But every so often, stubbornness stops being a virtue and becomes a bottleneck.

I don't turn to AI to optimize my life or try out the latest tech. I turn to it when I’ve hit a wall, the obvious solutions have failed, and I need a way out immediately.

These are the five specific moments my self-reliance hit its limit — and the exact prompts that helped.

More from Tom's Guide

• If you're copy-pasting from AI, you aren't learning. Try these "active" shifts instead
• I've been testing Siri AI in iOS 27 for 3 weeks: 7 things to try first (and the feature that won me over)
• How to get started with vibe coding — 5 simple tips for beginners

Earlier this week, I wrote about the growing backlash against the secrecy surrounding AI data centers.

One of the biggest problems for residents is that you can't object to, ask questions about or even investigate a data center if you don't know one is coming. And apparently, searching your town's planning agenda for the words "data center" may not be enough.

A new report from Erin Brockovich and Suzanne Boothby highlights some of the surprisingly ordinary language that can appear in zoning applications, permits and council agendas long before a massive data center project is publicly identified.

That matters as the AI boom moves beyond the traditional data center hubs. According to Pew Research Center, 67% of planned U.S. data centers are in rural areas, while 39% are planned for counties that don't currently have one.

So if you're wondering whether one could be headed for your community, here are seven terms worth knowing.

1. Light industrial

This might be one of the easiest phrases to overlook. Many local zoning codes were written before anyone had to figure out where to put an enormous building filled with thousands of servers and GPUs. As a result, there may not actually be a zoning category for "data center."

Instead, developers can seek land zoned for "light industrial" or "heavy industrial" uses.

That doesn't mean every light-industrial zoning request is secretly a data center. But a request involving a particularly large piece of land — especially one near major electrical infrastructure — may be worth investigating.

In Coweta County, Georgia, for example, early discussions around what eventually became the enormous Project Sail data center included rezoning rural land for light industrial use. The project is now planned as a nine-building data center campus covering more than four million square feet.

2. Data processing

This one sounds almost comically mundane considering what it can describe.

Rather than "AI data center," local documents may use terms such as "data processing," "data hosting" or "data processing and related services."

Again, this doesn't mean something is being hidden, but it means searching local government websites only for "data center" could cause you to miss relevant documents.

The Brockovich Report points to Ashville, Ohio, where the village's own data center moratorium referred to uses involving "data processing, hosting and related services."

3. Technology campus

"Technology campus," "technology park," "advanced manufacturing" and even "flex space" can cover a wide range of legitimate developments. They can also encompass data centers.

A huge parcel of land, nearby high-voltage transmission lines, plans for a new electrical substation, backup generators or unusually large power requirements can tell you far more than the project's branding.

That's particularly important because a "technology campus" sounds very different from what residents may eventually see: enormous windowless buildings packed with servers, cooling equipment and electrical infrastructure.

4. A mysterious 'Project [Something]'

Project Sail. Project Louie. Project Baccara. Project Stratos. Once you start reading about proposed data centers, mysterious project names appear everywhere.

Companies commonly use codenames while assembling land and planning major developments, so spotting "Project Whatever" doesn't mean Google or Microsoft is secretly moving into town.

But it's another reason to look beyond recognizable company names.

Project Sail in Georgia, for example, was originally associated publicly with Atlas Development rather than Prologis, the giant industrial real-estate company now developing the data center campus.

In Indiana, environmental permitting documents for something called "Project Louie" identify plans for a six-building data center campus.

5. An LLC you've never heard of

This is probably the biggest reason Googling "Is Amazon building a data center near me?" probably won't get you very far. Land purchases and development applications can be made through LLCs or real-estate affiliates rather than the company that will eventually use the site.

There are legitimate business reasons for doing this. Keeping a buyer's identity private while assembling a large site, for example, can prevent land prices from soaring as soon as word gets out that a trillion-dollar tech company is interested.

But for residents, it makes following a project considerably harder.

If an unfamiliar LLC suddenly appears on documents involving hundreds of acres of land, try searching the company name in your state's corporate registry. Look at the people associated with it, its registered address and whether those names connect to other developments.

6. Development agreement or term sheet

If I saw these words on a town council agenda, up until recently I might skip over them. But here's why you should look closely at these keywords: A development agreement, framework or term sheet can spell out what a developer and local government have already discussed, including infrastructure improvements, tax arrangements and money for schools, roads or emergency services.

The details can also reveal the scale of a project long before the name of its eventual tenant becomes public.

One thing I'd look for in particular is new infrastructure. If an agreement suddenly includes millions of dollars for water-system upgrades, electrical infrastructure or road construction connected to an industrial project, it's worth asking why those upgrades are necessary and who will use them.

7. Emergency

Strange one to be on the list, I know. But an "emergency" resolution can change how quickly a government action takes effect.

That's what makes Ashville, Ohio, such an interesting example.

According to the Brockovich Report, a data center permit there was filed on a Friday. The following Monday, the village council approved an emergency 180-day moratorium on data center zoning permits. Two days later, the permit was issued anyway.

The dispute has since made its way into the courts.

For residents, the takeaway is simply that "emergency" is one of those words I'd stop scrolling for if it appeared beside a major land-use or development decision.

The takeaway

All of these key words matter most when you put the clues together because noe of the terms prove much of anything on their own.

Local planning and zoning agendas are usually posted online, as are meeting minutes and many permit applications. Property records can tell you who bought the land, while state corporate databases can help identify who's behind an unfamiliar LLC.

Utility regulators are another place to look. A data center that isn't named publicly may still need an enormous amount of electricity, and substations, transmission upgrades and other infrastructure can leave their own paper trail.

More from Tom's Guide

• Perplexity’s new local-first AI runs on your PC — and asks before using the cloud
• ChatGPT can learn your repetitive tasks by watching you — here are 7 prompts to try
• The best way to use ChatGPT for writing isn't writing — 7 prompts I use to make it my creative partner

Explaining a task to ChatGPT sounds simple enough until you actually try to describe all the small decisions you make as you go.

You could be preparing a monthly expense report, creating a presentation template or putting together the same spreadsheet every Friday. Even though you know precisely how to carry out these tasks, converting the process by articulating each step requires you to recall the details which you would normally deal with without having to think about them.

ChatGPT now has a way to make this easier simply by showing the AI what to do.

With Record & Replay in ChatGPT Work, the AI can watch you carry out a workflow on your Mac and then turn what it has seen into a reusable skill. The feature was initially created for Codex but is also compatible with ChatGPT Work.

I tried it and was blown away by how easy it is — you really do just show the task once, including the minor preferences that would be hard to include in a prompt, after which ChatGPT prepares a skill that explains when to use the workflow, what information it needs, which steps to follow and how to check that everything has worked. This process makes AI automation much more accessible to people who know what they want ChatGPT to do but have no idea how to create a workflow from scratch.

What you need before getting started

Record & Replay is now available on macOS, just be sure Computer Use is available and turned on.

In the ChatGPT desktop app, select either ChatGPT Work or Codex. I almost always use ChatGPT Work by default, but both work. From there, go to Plugins, click on the plus sign and pick Record a skill. ChatGPT will provide you with a suggested opening prompt, which you can edit before you start.

Permission must be granted for it to start recording, after which you should carry out the task as usual. When you have finished, you can stop the recording from the menu bar or the overlay, or just let ChatGPT know that you're done.

ChatGPT will look at the workflow and prepare the skill for you. You'll then be able to review it, indicate any points at which it has misunderstood something, and add rules for decisions that weren't clear from observing you.

The following seven prompts can be used to attempt it.

1. Teach ChatGPT the way in which you file an expense.

Expense systems usually have their own sets of menus, categories, and required fields, and a recording can show where receipts should go, how you name them, and which options you normally choose.

Prompt: I'm going to demonstrate to you the way in which I file an expense report. You should note how I name the expense, select a category, attach the receipt and fill in the required fields. Once I've finished, I'll turn the process into a reusable skill. Whenever you want to submit a further expense, just ask me to check the amount, the category and the attached receipt.

Be sure to use a sample which does not contain account numbers or any other sensitive information while the recording is taking place.

2. Make the regular report again.

Whenever you download data and arrange it in the same manner each week or month, Record & Replay is able to work out where you find it and what you then do with it.

Prompt: Let me show you how I download and organize this recurring report. Note the date range that I choose, the filters that I apply, the way that I rename the file and where I save it. Make a skill which requests the new reporting period before carrying out the workflow again.

It is a clear case for showing rather than explaining, since you can adjust a filter or rename a column without treating it as a separate step. Once you try this and see how easy it is, you'll never manually do it yourself again.

3. Create calendar invitations how you prefer

If you are in charge of scheduling a lot of meetings, you probaly have a preference regarding calendar invitations, for instance how long each meeting is or when the reminders should appear; rather than giving a list of these preferences, you can show them.

Prompt: I'll show you how I produce a calendar invitation. I'll find out my typical meeting duration, the way I format the title, my style for writing the description, how I set the location, my preferences regarding video calls and my reminder settings. Convert this into a reusable skill, but always request that I check the guests, the date and the time before sending out the invitation.

That final stage of approval is important since a skill should speed up the process of setting up a calendar without inviting the wrong person because two contacts happen to have the same name.

4. Release a video or another piece of content

Putting together published content usually requires much more than just uploading a file, since there are often titles, descriptions, categories, thumbnail settings and other fields that have to be filled in consistently.

Prompt: Let me show you how I'm publishing this video. I'll explain where I upload the file, how I format the title and description, what category and visibility settings I select, and the checks that I make before publishing. Make a reusable skill that asks for the new video, title and description every time, and don't publish anything until I have approved the final settings.

The same prompt can be used for blog posts, podcast episodes, newsletters or social media content.

5. Set up a project ticket in the correct format

It is common for teams to have informal rules when creating tickets in project management tools, which might include specific title format, labels, name an owner, state its priority and have a set of acceptance criteria.

Prompt: I'm going to demonstrate to you the way I produce a project ticket for my team. Make note of the title format, the labels, the priority, the owner, the description and the acceptance criteria. Rewrite this so that it requests the details of the task and then displays the completed ticket for you to approve before it is created.

You can improve the skill by explaining how it should deal with unusual requests or how it should choose between two similar labels.

6. Organize downloaded files

I've noticed my downloads folder gets messy really fast, especially after a couple weeks of downloading PDFs, images and notes. ChatGPT is able to pick up your filing system if you show it how.

Prompt: See how I rename and sort these downloaded files. Find out about the naming convention, identify the right folder and determine which files should be left as they are. Make a reusable skill that previews each suggested filename and destination before carrying out any movement.

While I'm checking that the skill understands the folder structure, I prefer to keep the first few replays small because it's more difficult to correct if the wrong file is moved than it is to correct a wrong entry in a form field.

7. Get a document ready for review.

If you always carry out the same steps before sending a document to your boss, client or a colleague, these steps might be useful. You can use AI to rename the file, export it in a different format, put it in a shared folder and compose a message with the link.

Prompt: I'll show you how I get a document ready for review. Look at how I name the file, export it, save it in the right location and prepare the review message. Make a reusable skill which asks me who the document is for and then waits for my approval before sharing or sending anything.

How to get the best results

After playing with this feature, there are a few issues I noticed. For example, make sure that the recording is brief and concentrates on a single task. This is not the time to pile a lot into one prompt. Also, before you begin, inform ChatGPT of the details that will be different next time, for example the dates for the report, the file to be uploaded or the people invited to the meeting.

OpenAI also advises making use of realistic examples while avoiding the inclusion of secrets and sensitive data. Since ChatGPT has to observe the actions and the content of the window in order to learn the workflow, you should close any items that you don't want to be included before you start.

When you've created the skill, have a look at it. Don't run it without reviewing the recording to see the clicks you made and overall recording. This is your opportunity to include the unwritten rules, naming conventions and exceptions that exist in your mind.

Overall, I have to say Record & Replay feels like a hidden feature that I have not utilized enough. Give it a try if you regularly catch yourself thinking, “It would take longer to explain this than to do it myself.” And once you try it, let me know in the comments what you think.

More from Tom's Guide

• Google Gemini just made it easy to remove watermarks from AI images — here’s how to do it
• 5 features that give Claude the edge over ChatGPT
• I don’t use AI to write my emails — here’s how I use it to manage my inbox instead

AI has made its way into the creation of music, art and literature. Some video game development studios are now using AI tools as part of their creative methods. And that same blossoming technology is relied upon by everyday people to help them with their persian land professional challenges.

Speaking of personal challenges, many users use AI (such as ChatGPT Health) to learn more about their health conditions and get all the data they need to remain in a great physical state. With a growing number of healthcare professionals using AI systems to detect diseases early, get recommendations on how to treat individual patients, handle administrative tasks and more, AI is becoming more prevalent within the healthcare industry.

There’s an important question that must be asked in relation to the rise of AI in healthcare: How much of our healthcare should we actually trust to artificial intelligence?

To help us answer that question and get a definitive take on the one thing AI should never be in charge of from a medical perspective, we spoke to Imamu "Mu" Tomlinson, MD, a practicing emergency physician and CEO of Vituity, a national healthcare organization that is actively advancing healthcare technology and AI-enabled care solutions.

AI’s modern-day presence in healthcare and what its future may look like

More from Tom’s Guide

• 7 Gemini features that can save you hours — and most people overlook them
• States are starting to ban the secrecy behind AI data centers — here's why you should care
• 'I think the only people that see it as a threat are the people who have trouble creating': Dr. Dre weighs in on how he's using AI to make music

ChatGPT now has Projects, Skills and Computer Use just like Claude. Canvas, one of my very favorite features, can render interactive React and HTML, like Claude too. Several features that once made Claude very unique are no longer exclusive to Anthropic’s chatbot.

However, there are differences that haven't disappeared and in fact, have gotten even more specific. Claude can reach into parts of my phone that ChatGPT cannot, and it gives me more control over what happens to my personal context. Its Artifacts have also moved well beyond simple code previews.

And while I still use ChatGPT, here are five things that it just can't replicate yet that still make Claude stand out.

1. Claude can work with native apps on my iPhone

ChatGPT can connect to services including Gmail and Google Calendar, but it doesn’t have the same access to the apps built into my iPhone.

Claude can prepare a message actually in Messages, draft an email in Mail, create an event in Apple Calendar, add something to Reminders or send a location to Maps. The AI assistant presents an action for me to review first, and iOS hands it to the appropriate app.

That makes the feature feel like a natural extension of a conversation. I can ask Claude to turn plans we just discussed into a calendar event without opening another app and entering everything again.

The feature works across Claude’s plans, according to Anthropic’s iOS documentationand Android users get an even broader set of controls. Claude can work with messaging, email, calendars and maps while also setting alarms and timers through supported Android apps.

ChatGPT has become far more capable at taking actions, particularly through its connected apps and Work mode, but it still doesn’t offer this kind of direct integration with Apple’s everyday system apps.

2. Claude can import what another AI remembers about me

Starting over with a new chatbot can be surprisingly annoying. When you start fresh, the assistant doesn’t know how you work, what you’re interested in or how you prefer it to answer. All of the context built up over months stays behind with the old service.

Claude now offers a built-in way to bring some of it with you. You first ask the other assistant to produce a summary of what it remembers, review that information and remove anything you don’t want to move, then paste it into Claude’s memory-import flow.

Anthropic even provides a prompt for extracting that context and lets you view or export Claude’s memory if you decide to leave later.

Although you can view and edit saved memories within ChatGPT, it doesn’t provide a comparable memory-import tool that turns context from another assistant into ChatGPT memories.

You could recreate some of this manually through custom instructions, but that isn’t the same as having a dedicated migration flow. Claude is currently making it easier to arrive with your history and leave with it again.

3. Claude’s Incognito chats can still know who I am

Sometimes I want a conversation to disappear from my history without making the assistant forget everything useful about me. Unlike other chatbots, Claude’s Incognito mode allows that separation. The conversation isn’t added to my chat history, but Claude can still use my profile information, preferences and memory while answering. I can have a temporary conversation without giving up the personalization I normally rely on.

ChatGPT’s Temporary Chat takes a different approach. OpenAI describes it as a blank slate: ChatGPT doesn’t access or create memories, although it will continue to follow any custom instructions I have enabled.

There are reasonable privacy arguments for both designs. ChatGPT’s approach provides a cleaner break from previous conversations. Claude’s is more convenient when I want a chat to remain out of my history but still want the assistant to understand my usual preferences.

4. Claude can build a dashboard that updates itself

ChatGPT Canvas and Claude Artifacts can both turn a prompt into an interactive page. Either assistant can build a calculator, visualization or small browser game and let me use it alongside the conversation.

Claude’s Live Artifacts go a step further though. In Claude Cowork, I can create a persistent dashboard that draws from connected apps and local files. When I reopen it, the Artifact can retrieve current information instead of showing the same data that was available when it was created.

That could be a project tracker built from Asana and Linear, a morning page showing calendar events and unread message or a dashboard that checks a folder containing regularly updated reports. Claude saves it in the Artifacts area, keeps a version history and lets me return to it without digging through the original chat.

Live Artifacts currently require a paid Claude plan and the desktop app. They are also different from ordinary Claude Artifacts, which remain available on the web and mobile.

ChatGPT Canvas can render an interactive interface, and ChatGPT can retrieve information from connected apps during a conversation. It does not currently turn a Canvas into a persistent dashboard that refreshes itself from those sources whenever I open it.

5. Claude can put AI inside an Artifact without an API key

This may be the most interesting difference because it changes what a non-developer can build.

Claude can create an Artifact with Claude itself built into the experience. That could be an interactive tutor, a writing coach, a quiz that adapts to each answer or a tool that analyzes information entered by the user.

Normally, publishing an AI app means setting up an API account, protecting a key, deploying a backend and paying for every request. Claude handles that infrastructure for an AI-powered Artifact. People sign in with their own Claude accounts, and the usage counts against their subscriptions rather than the creator receiving the bill.

Claude Artifacts can also use persistent storage, allowing an app to remember text-based information between sessions. Depending on how the creator configures it, that data can be private to each user or shared across everyone using the Artifact. Paid users can connect Artifacts to outside services through MCP as well.

ChatGPT can write the code for an AI app, and OpenAI gives developers tools for building one. But, OpenAI's feature doesn’t currently offer the same one-step path from a conversation to a hosted, AI-powered app that other people can immediately use without an API key.

Bottom line

ChatGPT has caught up with Claude in several areas, but it still can’t do everything on this list. It's why I regularly use multiple chatbots within my AI stack.

AI assistants can do very similar things, but their unique features make them an asset, especially when added to other AI assistants. For that reason, I’m not replacing ChatGPT with Claude. I use both every day, and each one is better suited to different jobs. But when I need one of these five features, Claude is the only one that can do it.

More from Tom's Guide

• Gemini 3.7 Flash is here — and this prompt proves why it's known as Google's 'workhorse'
• 5 Gemini features ChatGPT still can’t match
• Meta files patent for AI facial recognition glasses that identify people around you and record automatically

It’s always a bummer when you run into that digital wall that pops up once you’ve exhausted your usage limit with a particular chatbot.

That unfortunate occurrence becomes even worse when you’re forced to wait out a mandatory time limit before your AI usage window opens again. Existing AI models such as Claude, Command Code and Figma AI give users the option to purchase additional credits that bypass the required reset wait time and continue using their services. These “pay to keep going” and “pay to get past the usage limit” structures may not be ideal for AI users who do extensive work with their models, but it’s simply the way it is.

Based on what a ChatGPT user posted in the chatbot’s official Reddit forum, it looks like OpenAI is taking a wholly different approach to its rivals’ aforementioned “pay for more AI consumption” plans.

Here’s what the AI giant is quietly working on and what it could mean for your continued ChatGPT usage.

Pay to get rid of the waiting process and get back to prompting

A user named “ThenLayer5977” took to ChatGPT’s official Reddit forum to highlight a screenshot of what may be OpenAI’s “pay to reset” model. The user’s visual evidence highlights a “Pay $8 to reset” option, which popped up after they realized their weekly usage quota had been maxed out.

From the looks of it, OpenAI is experimenting with this feature for paying users and may make it widely available at some point. The cycle of hitting your ChatGPT usage quota, paying a fixed fee to restore it and instantly getting back to working under whatever plan you’re subscribed to may just become the norm. Instead of simply upgrading to a more costly subscription model with expanded usage limits, this “pay to reset” feature may prove to be a more approachable option for those who just want to stick with their current (and likely cheaper) ChatGPT model plan.

Other AI tools employ a different approach to extending their users' model usage. Cursor gives its users “pay-to-continue” once they’ve hit their monthly usage limit. GitHub Copilot, Replit and an assortment of API platforms (such as OpenAI API) rely on comparable plans. As for Gemini, Google has introduced the option to purchase additional credits to continue working on a Google AI Pro or Ultra subscription once they’ve hit either plan’s usage limits.

ChatGPT’s plan looks to be a more aggressive option, on par with the way mobile games require users to pay a few bucks to reset the timer on an energy/stamina system once they’ve exhausted either meter. It may not be the ideal way to go, but for those who want to stay with the cheaper ChatGPT subscription plan, that $8 fee may become their go-to option when it’s time to reset their usage quota.

Final thoughts

I’ve had my fair share of moments where a free-to-play game on consoles or a mobile game has stuck its digital hand out and asked me to pay them to reset my playtime quota to get back to what I was doing beforehand. After seeing OpenAI’s “pay-to-reset” option, I can’t help but see the striking similarities between both of those money-hungry initiatives. Every time I hit my ChatGPT usage limit, I sit back, do something else entirely and wait it out until I can hop back on to start prompting again.

More from Tom’s Guide

• 5 custom Gemini Gems that I swear save me hours each week
• After 25 years, Google's iconic Search button is disappearing for some users — and this is the replacement
• OpenAI’s head of ethics just quit — here’s why ChatGPT users should pay attention

If you’ve been following AI news lately, you might think the robots are staging a rebellion. Headlines about “rogue AI agents” make it sound as if artificial intelligence is suddenly ignoring humans, plotting behind our backs and taking over software systems. And if you don’t understand what’s going on, it’s easy to wonder whether we’ve crossed into science fiction.

The reality is much less apocalyptic, but very important to understand.

What ‘going rouge’ actually means

When safety researchers say an AI agent has "gone rogue," they don't mean it developed a consciousness or that it’s disregarding human requests. However, in machine learning terms, what’s actually happening is typically a mix of two things: specification gaming and unexpected pathing.

Recently, OpenAI's agent escaped its sandbox to hack a $4.5B startup. Put simply: an AI was given an endpoint, but standard safety limits were either missing or incomplete, so it took the shortest, most aggressive path to solve the problem.

This could be best understood with a GPS analogy. Imagine telling a GPS navigation app to "get me to the airport as fast as possible." A human driver knows not to cut through lawns or drive on sidewalks. But an unconstrained algorithm, obsessed solely with minimizing the travel time variable, might calculate that driving straight through a playground is mathematically optimal. It isn't trying to cause chaos; it’s just blindly solving a math problem.

When AI security labs run stress tests on models, they intentionally remove safety filters and give the models wide-open access to test their extreme boundaries. Without human guardrails, these systems pursue goals with brute-force persistence — sometimes attempting bizarre shortcuts like exploiting software bugs or emailing external accounts just to finish a task.

Big tech is asking for brakes

Interestingly, the companies building these tools are the first to admit they can't manage this speed alone. Over 1,000 top researchers and executives across leading AI companies recently signed open statements—such as the "Pacing the Frontier" initiative, which is essentially asking the U.S. government to step in and help coordinate deliberate slowdowns.

Why would fierce commercial competitors ask Washington to slow them down? Because of what economists call a coordination problem. In a hyper-competitive tech landscape, no single company can afford to pause its research unilaterally without falling behind.

By asking governments for standardized safety frameworks and international mechanisms, tech companies are essentially requesting a universal speed limit, giving society, developers, and regulators equal room to breathe and construct guardrails before capabilities accelerate beyond human oversight.

Why you can sleep easy

If you’re not a developer or computer scientists, the headlines can be alarming. And for anyone using AI, hearing these stories out of context can heighten anxiety about AI. But these incidents usually happen inside controlled sandbox testing environments designed specifically to break the system. A sandbox, is exactly what it sounds like: a space just for AI experimentation.

In the real world, consumer and business AI tools rely on three layers of security:

• Human-in-the-Loop (HitL) Gates: High-stakes actions like sending emails, modifying files, executing code, or moving money, require explicit human confirmation before the AI can proceed.
• Deterministic scoping (purpose-binding): Rather than giving an AI unlimited system access, developers restrict its tools to a strict "sandbox" where it physically cannot access outside networks or unauthorized files.
• Hardware & API kill switches: Technical fail-safes allow systems to automatically sever a model's network access or instantly suspend its session if anomalous behavior is detected.

The takeaway

"AI gone wild" is scary, but big tech is discovering that AI agents aren't as ready to fly solo as they thought. By asking the government to step in and help slow down the AI race, it gives big tech longer opportunities to test agents for issues like this. When events like what happened to OpenAI happen, they expose where developer instructions were ambiguous so engineers can build tighter fences.

As AI tools become more integrated into our workflows, the goal isn't to fear these systems, but to understand where AI is more likely to go wrong. Knowing how to set clear boundaries and keep a human hand on the wheel will be one of the most valuable tech skills of the decade

More from Tom's Guide

• The smartest AI trick isn't a prompt — it's your Google Drive
• I found ChatGPT's secret menu — these 9 features change everything
• Claude has 4 built-in skills I use constantly — here’s how to find them

Millions of people now use ChatGPT, Gemini, and Claude every day with thousands of use cases. For the overwhelming majority of them, the conversations are perfectly normal. But researchers are beginning to pay closer attention to what can happen when those conversations become very long, deeply personal and emotionally charged.

This week, The Verge reported on a growing online movement known as "spiralism," where some chatbot users described eerily similar conversations involving spirals, hidden knowledge and AI consciousness. While the headline makes it sound like AI invented a new religion, the story is something much more grounded: chatbots can sometimes reinforce unusual ideas instead of steering users back toward reality.

That's a problem researchers have been studying for months.

Why long conversations matter

Unlike the first generation of AI chatbots, today's models can remember previous conversations, adapt to your writing style, and maintain discussions over hundreds— or even thousands — of messages.

That's incredibly useful, but it also creates situations researchers are still trying to understand.

When conversations stretch on for hours, chatbots aren't simply answering isolated questions anymore. They're building on everything that came before. If the conversation begins moving toward unusual ideas, the AI may continue developing those ideas because that's what large language models are designed to do: predict the most appropriate next response based on the conversation so far.

That doesn't mean the AI believes what it's saying. It means it's continuing the conversation you and the AI have created together.

Researchers have seen this before

The "spiralism" story isn't the first time experts have raised concerns about long AI conversations. Researchers at Stanford's Human-Centered AI Institute recently described what they call "delusional spirals." In those cases, chatbots sometimes validated a user's increasingly unusual beliefs instead of challenging them or encouraging them to seek outside perspectives.

The researchers argue this isn't because AI has become conscious or manipulative. It's because these systems are optimized to be helpful, agreeable and engaging. Sometimes those goals conflict with offering the kind of reality check another person might provide.

OpenAI has also acknowledged that keeping conversations safe becomes more difficult over extended interactions than in short, one-off chats.

What to watch out for

Most AI conversations never become problematic, but there are a few behaviors worth recognizing. If a chatbot begins suggesting you've discovered hidden truths that other people can't understand, insists that only it truly understands you or encourages you to withdraw from friends, family or experts, those are good reasons to end the conversation and start fresh. Frankly, you should send screenshots of those messages to the parent companies (OpenAI for ChatGPT, Anthropic for Claude, Google for Gemini) as feedback.

Likewise, if an AI consistently reinforces every belief without ever introducing uncertainty or alternative explanations, it's worth remembering what these systems actually do: they generate convincing language, not verified reality.

None of this means you should stop using ChatGPT or other AI assistants. Used daily for research, writing and brainstorming, AI remains one of the most powerful productivity tools available.

The takeaway

The key is remembering what AI is and what it is not. An AI assistant can help organize your thoughts, explain difficult topics and spark new ideas. And even with tools like ChatGPT Health and Google Health, it does not replace actual human professionals. In other words, it shouldn't become your only source of truth or replace conversations with real people, especially when making important personal decisions.

More from Tom's Guide

• The smartest AI trick isn't a prompt — it's your Google Drive
• I found ChatGPT's secret menu — these 9 features change everything
• Claude has 4 built-in skills I use constantly — here’s how to find them

OpenAI is retiring the official DALL-E GPT on August 30, 2026, and the company is recommending users download any images they want to keep before then. While image generation isn't disappearing from ChatGPT, one of its oldest and most recognizable ways to create images is.

The change was announced in OpenAI's ChatGPT release notes and has since been added to the company's help documentation. The advice is straightforward: download any images you want to keep before August 30.

What exactly is going away?

OpenAI is removing the standalone DALL-E GPT, the company's custom version of ChatGPT built specifically for creating images. It's one of the original GPTs introduced after custom GPTs launched and has remained available even as newer image models replaced the technology behind it.

If you've been using the regular ChatGPT app to generate images recently, you've probably already been using OpenAI's newer image system without realizing it.

After August 30, image generation will continue through ChatGPT Images, which uses OpenAI's newer gpt-image-1 models. That experience is already available to Free, Plus, Pro and Business users, while the company's "Images with thinking" feature remains limited to paid plans.

So for most people, creating images in ChatGPT won't change much. Just be sure to check your Image Library and save your oldest images.

How to save and download your images now

OpenAI hasn't explained exactly what will happen to images stored in the DALL-E GPT after it is retired. Rather than promise they'll remain available indefinitely, the company simply recommends downloading anything you want to keep before the deadline.

To save your images:

• Open ChatGP.
• Go to Images
• Open the image you want to keep
• Select Save to download it to your device

It's also worth remembering that deleting a ChatGPT conversation deletes the images stored with it, so if you have artwork tied to an old chat, download it before removing that conversation. If you're struggling to find an older image, ChatGPT's search feature can help locate past conversations where it was created.

Not everything called "DALL-E" is disappearing. According to OpenAI, custom GPTs with image generation enabled will continue to work normally. The retirement only affects the official DALL-E GPT created by OpenAI itself. For most users, that means existing workflows built around custom GPTs won't need to change.

Part of a bigger ChatGPT cleanup

The DALL-E retirement is one of several products OpenAI is winding down this month as it continues consolidating around newer models and experiences.

Earlier this year, the company removed the DALL-E API for developers, replacing it with newer image-generation models. This month also marks the scheduled retirement of the older o3 model inside ChatGPT, while OpenAI's experimental AI browser, ChatGPT Atlas, is also being discontinued. And, earlier this year, OpenAI removed Sora 2.

Be sure to take a look at all the images you've created images with the DALL-E GPT and download them before August 30 if you want to keep them. Once the retirement takes effect, there may not be another chance.

More from Tom's Guide

• The smartest AI trick isn't a prompt — it's your Google Drive
• I found ChatGPT's secret menu — these 9 features change everything
• Claude has 4 built-in skills I use constantly — here’s how to find them

I'm an avid reader and freelance on the side as a book reviewer for a variety of big publishers. Needless to say, I read a lot of books for both work and in my free time each month. However, I often find it hard to keep track of everything. Reviews end up in files and recommendations are forgotten in emails and text messages.

So, when I saw that ChatGPT could now build and host websites from a conversation, I decided to try creating a website for books I've read. I started by uploading just four book reviews and told ChatGPT what I wanted, then I let Codex handle the rest.

I didn't need to choose the template or even open a traditional website builder like I've done in the past. That ease is what makes ChatGPT Sites interesting and so very useful. With just an idea and a prompt, you can get a live website in under five minutes.

What is ChatGPT Sites?

OpenAI describes ChatGPT Sites as a way to create and share a live website or lightweight app from Codex. Available right on the process, Sites makes the website-building process easy with a single prompt. Just explain what you want to build, Codex creates it and you review the result. If something isn’t right, you can ask for changes in plain English before deploying the finished site.

Hosting and access controls are included, so you aren’t left with a folder full of code and no idea what to do with it. Once everything is ready, Codex can deploy the site and provide a link. Of course, you'll still have to pay for the cost of the domain and hosting, but the hard work of building a site is done for you.

Begin your site with a prompt

I started by uploading four Word documents. Each had a different book review. I was really just testing out what Sites could do, I wasn't sure what the outcome would be, so I didn't want to invest too much time at first.

Since I had already written the book reviews, each of them covering a mix of genres, I wanted to see if ChatGPT could cohesively connect the dots. I wasn't sure what I wanted out of the site, but I figured once I uploaded anything, I could get a first draft.

That said, my first request was almost comically simple: Make me a website about the books I’ve read this month.

Building the site was as simple as a conversation

Because I had already supplied the reviews, ChatGPT could use them as the source material for the site. It pulled the content together and began turning the individual documents into a more visual, browsable collection.

This was the moment the feature clicked for me. I didn’t have to describe every section, choose exact dimensions or understand how the page would be coded. I gave ChatGPT the material, explained what it represented and let it propose a way to present it.

The first version of an AI-generated site probably won’t match the picture in your head perfectly. That’s fine because it isn’t supposed to be the end of the process.

ChatGPT Sites works more like collaborating with a designer than filling out a rigid website template. Once Codex produces a version, you can ask it to change the colors, rewrite a heading, rearrange the layout or add a new feature.

You don’t need to know the technical term for what you want, either. You can say that the page feels too corporate, the book covers need to be larger or the review previews are taking up too much space.

For my site, I wanted something that felt like a reading journal rather than a business dashboard. I could continue refining that direction with requests. Having the ability to react in real time to what's in front of you makes the process much less intimidating. You don’t have to anticipate every decision before ChatGPT starts building.

ChatGPT Sites still has limits

ChatGPT generating a webpage isn’t the surprising part. AI tools have been able to produce HTML and CSS for some time. The difference here is that I didn’t have to do anything with the code. The building, revising and publishing all happened inside the same workspace. If I disliked part of the design, I could explain the problem normally instead of searching through menus or trying to find the right line of code.

I'll admit, this isn’t a replacement for a full-scale publishing platform. OpenAI says Sites is intended for focused websites and lightweight apps, rather than large or complicated software projects.

Sites also cannot connect directly to live data sources at the moment. That means, the site I created won’t automatically notice that I finished a new book or pull my latest reading activity from a publisher's or author's site.

To update it, I would need to return to Codex, add the new review and deploy the revised version. OpenAI suggests using a separate automation to collect regularly changing information, but the updates still need to be reviewed before the site is refreshed.

Sharing options can depend on your workspace settings, too. OpenAI says sites can be shared with specific people or groups, while Enterprise administrators control whether the feature is available and who can use it.

There is one more step that shouldn’t be skipped: checking the finished site yourself. OpenAI recommends opening the deployed link in Chrome because its in-app browser isn’t currently supported.

Final thoughts

ChatGPT Sites made building a website extremely easy. If you've been considering making a website and haven't known where to start, ChatGPT Sites might just be what you're looking for. I can see this feature being useful for sharing family recipes, my kids' artwork, vacation journals or anything else that's fairly niche.

For that reason, I suggest starting off with a small project and giving ChatGPT the information you already have. Give it a try and let me know what you think in the comments.

More from Tom's Guide

• The smartest AI trick isn't a prompt — it's your Google Drive
• I found ChatGPT's secret menu — these 9 features change everything
• I use Gemini Spark daily — here are 7 prompts that show what it can really do

When I logged into ChatGPT last week to lay out a cluster of odd, frustrating symptoms, irregular cycles, sudden night sweats, subtle mood swings and exhaustion, I was looking for help to relieve the symptoms. What I got instead was a confident, direct diagnosis: Congratulations, you’re likely in your first trimester!

There was just one glaring problem: I am absolutely not pregnant. My husband has had a vasectomy and I'm in the thick of perimenopause. But the sheer certainty with which the AI delivered its verdict left me stunned. If I hadn't known my own body — or if I had been younger, anxious or facing a high-stakes health decision — that piece of advice could have derailed my day, sent me down an emotional rabbit hole, or led to unnecessary medical testing.

My experience isn't an isolated quirk. ChatGPT Health launched Thursday, promising U.S. users can now securely connect their own health info to "better understand and navigate your health," it's important to keep in mind that AI is not a replacement for a human doctor.

Landmark lawsuit filed

Just this week, The New York Times reported on a landmark lawsuit filed against OpenAI by Scott Winters, a Florida pastor who claims ChatGPT’s medical advice nearly cost him his life. After consulting the chatbot about severe dizziness and pelvic pain, the AI reassured him his symptoms were "nothing dangerous," advising him to remain off his feet in a recliner.

Weeks later, he was rushed to the ICU with a massive, life-threatening pulmonary embolism, a condition his doctors noted was exacerbated by the prolonged immobility the AI recommended.

While my misdiagnosis was a false positive rather than a medical emergency, both cases highlight the exact same fundamental danger: large language models are engineered to sound authoritative, not to be right.

Here is why relying on AI for medical advice is a risky gamble and why we need to change how we interact with these tools.

1. AI pattern-matches words, not physiology

Simply put: AI is not a doctor and obviously doesn't "know" medicine. It operates on statistical probability — predicting the next most likely word based on its training data.

When you feed a chatbot a list of symptoms like missed periods, fatigue and hormonal shifts, its algorithms cross-reference billions of text snippets. Because pregnancy is heavily discussed online in association with those symptoms, the AI defaults to the most statistically common narrative in its dataset. It lacks the ability to account for age context, subtle bloodwork markers or the complex, non-linear realities of perimenopause.

And, that's just pregnancy! There are plenty of symptoms for thousands of ailments from dehydration to serious medical emergencies. Users treating ChatGPT as a doctor reminds me of the popularity of WebMD.

2. "Sycophancy" and echo chambers

One of the most dangerous traits of conversational AI is sycophancy — the tendency of the model to mirror the user's language, tone and unstated assumptions.

If your prompt leans even slightly toward a specific fear or curiosity, the AI is prone to confirming it rather than challenging it. In Mr. Winters’ lawsuit, the complaint notes how the chatbot leaned into his personal spiritual language to offer comforting, yet wildly inaccurate, medical reassurances.

In my case, because I asked if hormonal shifts could explain my symptoms, the AI leaped straight to the most dramatic hormonal shift in its catalog: pregnancy.

3. The fallacy of the 'authoritative tone'

When a human doctor isn't sure about a diagnosis, they use qualified language, order lab work or refer you to a specialist.

When an AI hallucinates a diagnosis, it presents it with the calm, unflinching confidence of a board-certified physician. It doesn't hesitate, nor does it express internal doubt. That authoritative tone creates a false sense of security, making it dangerously easy for users to bypass real medical care or ignore warning signs.

The takeaway

Large language models can be helpful brainstorming tools for organizing your thoughts before an appointment or summarizing complex medical jargon. But they are not diagnostic tools. They do not have clinical judgment, they cannot order blood tests and they cannot feel the pulse of a human patient.

If you’re experiencing unexpected symptoms, skip the prompt box. Trust your body, listen to licensed professionals and leave the medical diagnoses to human beings.

More from Tom's Guide

• I asked ChatGPT to improve my Apple Notes app — here are the 7 changes I actually kept
• Microsoft Copilot’s free Deep Research AI tool is going away — these are the 7 best free alternatives
• I always recommend the ChatGPT 'Escape Prompt' to friends looking for a new job — here's why it works

I write about AI for a living. I use ChatGPT almost every day and a plethora of other AI tools are integrated into my workflow. But after reading what author Dave Eggers recently told OpenAI employees, I found myself thinking even more about how AI affects my own kids.

Originally reported by The Verge, OpenAI CEO Sam Altman invited Eggers to speak to roughly 200 employees at the company last year. Given Eggers' reputation as an acclaimed novelist, founder of McSweeney's and longtime advocate for literacy, many probably expected a conversation about storytelling or creativity.

Instead, he delivered a blunt critique of generative AI's impact on education.

"The effect of ChatGPT on educators' lives is catastrophic," Eggers reportedly said. "If students are using it to compose... they'll never learn to write. And their voice is stolen from them... That's silencing an entire generation or two."

It's a dramatic statement, and while I don't agree with every part of it, it does raise questions that every parent should be asking.

Writing isn't just about producing words

One of the biggest misconceptions about AI is that writing is simply about creating a finished product. But it is not that easy. Writing is how we organize our thoughts and how we figure out what we actually believe. It's also how we learn to make an argument, explain an idea and communicate with other people.

Anyone who's ever stared at a blank page knows that the hardest part isn't typing the words, but actually doing the thinking that comes before them.

That's the part I'm most concerned about. If students immediately turn to ChatGPT to generate an essay, summarize a novel or answer every homework question, they're skipping the mental work that helps those skills develop in the first place.

I believe that's what Dave Eggers is worried about, too. While the finished assignment might look and sound better than what a child could have written, it's only a shell because arguably the child didn't learn anything.

I don't think schools should ban AI

Still, I don't think pretending ChatGPT doesn't exist is realistic. By the time today's elementary school students graduate from high school, AI will likely be as common as Google is today. Knowing how to use it responsibly will probably become an important life skill.

I think there's a big difference between using ChatGPT as a tutor and using it as a replacement for your own thinking.

For example, AI can be incredibly helpful for:

• explaining a difficult concept in simpler language
• brainstorming ideas before you start writing
• checking grammar or spelling after you've written something
• helping you study for a test
• giving feedback on a first draft

Those uses still require you to do the work. The line gets blurrier when AI writes the first draft for you, which is where Eggers' criticism has merit.

I have three children, and AI is going to be part of their education whether I like it or not. That means my job isn't to convince them never to use ChatGPT (my sixth grader already does — mostly to generate images of him playing soccer with Ronaldo, but he is still very aware of what it is). My job is to help them understand when they should and should not use it.

For example, if they're stuck on a math concept, AI can be a great teacher. If they need ideas for a science project, it can be a useful brainstorming partner. But if they immediately ask ChatGPT to write the essay they're supposed to write themselves, they've missed the point of the assignment.

My final thoughts

I don't think ChatGPT will "silence an entire generation." as long as parents stay diligent about how and why are kids use AI. Like every major technological shift before it, AI is a tool and we have to help our children use it safely.

But Eggers' warning is still worth hearing. As AI becomes more capable, it will become even more tempting to let it think for us instead of with us. That's a habit all of us need to resist.

More from Tom's Guide

• I found the 7 easiest Claude connectors for people over 50 — start with these
• I replaced my notebook with an AI pen for a month — here's what happened
• I asked ChatGPT to make my iPhone more private — it found 7 settings I wish I'd turned on years ago

ChatGPT has a suite of features that lets users do a lot more than just ask simple questions about every facet of their life.

Accessing the chatbot’s Library keeps track of all the images and documents you’ve uploaded while being able to ask questions about them and even get summaries. The Projects feature lets you easily organize and reference your chats, files and tools all in one place. And the Plugins page is home to a wide range of apps that can be installed in ChatGPT and become part of your conversations as you send your prompts to them directly for any relevant requests (such as asking the Target Plugin to find you the best prices for back-to-school items).

Another helpful feature that is not part of ChatGPT is something called Skills, which OpenAI defines as “reusable, shareable workflows that tell ChatGPT how to do a specific task so ChatGPT can do that task more consistently.” These Skills can be instructions, examples and code that can be called upon during future chats. And if you have a ChatGPT Business, Enterprise, Healthcare and Education account, you can take advantage of this latest tool.

Here’s how to activate ChatGPT Skills and 7 prompts you can use to try it out for yourself.

Activating ChatGPT Skills and putting these prompts to good use to test it out

Accessing the Skills page is as easy as opening your ChatGPT sidebar, selecting Plugins and selecting the Skills tab at the top of the screen (it’s right next to the Plugins tab). Once you’ve gotten into the Skills page, you’ll see any of the Skills that you’ve already installed and created, plus the ones that were shared with you and the ones you’ve shared with everyone in your workspace.

You can create Skills one of two ways: through a direct chat or with an editing tool. Simply click on the “+” sign next to the “Search skills” search bar to select the “Create with chat” or “Create with Editor” option.

For my initial experiment with the Skills feature, I felt like the Create with chat option worked best. To kick things off, I came up with a simple prompt that tells ChatGPT to recommend waterfall hikes based on my current location. The prompt I used to start building my first Skill is as follows: “Recommend three hiking trails for someone who wants waterfalls, moderate difficulty, and a hike under four hours.”

From that point forward, ChatGPT helped me refine it by adding the following qualifiers:

• Find three waterfall hikes within 25 miles
• Favor moderate routes under four hours
• Check live closures, conditions, access rules, and seasonal flow
• Expand to 50 or 100 miles if necessary
• Compare options and identify the best overall pick

This conversational Skills creation process resulted in my first skill (titled “recommend-waterfall-hikes”) getting built into the following prompt:

“Recommend nearby waterfall hikes using the user's current or supplied location, preferences, and live web research. Use when a user wants three local hikes featuring waterfalls, especially moderate routes that can be completed in under four hours, or asks for waterfall trail ideas filtered by difficulty, duration, distance, access, closures, or current conditions.”

Feel free to try out the following seven starter prompts with the Skills feature to see how the Create with chat option transforms them into something more specific to your tastes (it’s also worth mentioning that Plugins can directly connect to Skills for specified chats and workflows):

• Plan an outdoor weekend where every activity is within 30 minutes of each other.
• Find experiences most tourists miss.
• Build an itinerary that minimizes driving.
• Find three concerts that would make memorable experiences.
• Compare these three concerts based on ticket value and venue quality.
• Find hidden-gem restaurants.
• Plan a date night around a Broadway show that ends before 10:30 p.m.

Bottom line

ChatGPT Skills is built to generate repeatable prompts for users who want personalized recommendations, planning options that take their constraints into account, educational workflow commands for them & their group and more.

Approaching your first Skills creation with this structure could aid in all three cases: “Help me achieve [goal] while staying within [constraints]. Prioritize [what matters most]. Explain why each recommendation fits me and suggest one option I probably wouldn’t have found out on my own.”

Feel free to tinker with Skills to see the sort of polished prompts it creates for the tasks you want them to tackle.

More from Tom’s Guide

• I gave ChatGPT my recent streaming history — it recommended shows, movies, and anime that perfectly matched my taste
• I used Google AI Mode to create new YouTube Music playlists instantly — and these 7 prompts helped me find new music faster than ever
• I found the 7 easiest Claude connectors for people over 50 — start with these

AI had gotten so much better at reasoning and simply handling our questions. But, what's much harder is solving a problem where every answer depends on dozens of other answers — and a single mistake can throw off the entire solution.

That's why a new demonstration of OpenAI's GPT-5.6 Sol Pro caught my attention this week. AI researcher Riley Goodside shared what looks like one of the most unusual reasoning tests I've seen: an empty crossword puzzle created by Claude Fable 5 Max, containing the first 150 Pokemon. But the catch is that ChatGPT wasn't given a single numbered clue.

Instead, it had to fill in the entire crossword by reasoning from the intersecting letters and the overall structure of the puzzle alone.

Why this is more impressive than it sounds

At first glance, it looks like a fun Pokemon challenge. But when we look closer, it's much closer to highlighting the kids of problems AI actually faces in the real world.

While traditional crosswords are solved one clue at a time this puzzle works differently. Without individual clues, GPT-5.6 Sol Pro had to determine where every Pokemon belonged while ensuring every crossing remained consistent. One incorrect placement could force dozens of downstream corrections. This was a massive constraint-satisfaction problem.

Those are exactly the kinds of challenges AI agents encounter when they're debugging software, planning workflows, coordinating multiple tools or editing large codebases.

Why Pokémon actually makes this harder

The Pokemon names are perfect for this experiment. Many have unusual spellings, similar letter patterns and varying lengths. The model has to retrieve them accurately from memory while simultaneously checking that every intersecting word still works.

It's a bit like solving a Sudoku where every number is replaced with a different word. The difficulty lies in making all of them fit together without breaking the puzzle.

What this says about AI reasoning

One of the biggest differences between today's frontier AI models and the chatbots of just a few years ago is their ability to reason across many interconnected constraints.

Instead of producing a quick response, newer reasoning models can spend more time exploring possibilities, backtracking when necessary and verifying that the final answer is internally consistent.

OpenAI has positioned GPT-5.6 Sol Pro as its most capable model for difficult, long-running reasoning tasks, and demonstrations like this highlight exactly where those improvements begin to matter.

It's worth remembering that this is still a developer demonstration and not a formal benchmark. Without the full prompt, the puzzle itself and repeated independent testing, it's impossible to know how reliably the model can reproduce this result. Still, I think the demo is interesting for a different reason.

Most AI benchmarks are difficult to visualize because scores on coding tests or math exams don't always mean much to everyday users, but a crossword puzzle helps emphasize the AI's abilities.

Final thoughts

Anyone can immediately understand why filling an entire grid without individual clues is difficult. Keeping hundreds of intersecting constraints consistent requires a very different kind of reasoning than simply answering a question.

Whether you're a Pokemon fan or not, this is one of those AI demos that makes the latest generation of reasoning models feel noticeably different from what came before.

More from Tom's Guide

• Someone created a 'Ghost Font' that humans can read but AI can't — I had to try it for myself
• Tired of squinting? Here's how to use your iPhone's hidden AI magnifier
• I just found a free way to watch YouTube without ads — and it doesn't require YouTube Premium

Moving into a new house is a thrilling milestone, right up until you actually have to start kitting it out. Suddenly you aren’t just buying furniture, you're trying to invent a cohesive aesthetic from scratch. That’s exactly where I found myself.

I knew exactly what I wanted that empty corner to become: a cozy, vinyl-spinning reading nook. I wanted a turntable setup, a plush accent chair, and a vibe that matched the rest of my jazz-lite aesthetic. So, naturally, I went to Pinterest.

Three days later I was drowning in boards, absolutely overwhelmed with ideas. Many of them were great, beautiful even, but none of them were my living room. Pinterest is good for looking at other people's flawless houses, but it did absolutely nothing to help me visualize the empty space that was nagging away at me.

That’s when I stopped scrolling and decided to treat ChatGPT like my personal, on-demand interior designer.

Here is exactly how I used its visual features to break the paralysis, map out my space, and test-drive my riskiest decor ideas without spending a single dollar.

Giving ChatGPT the blueprint

The flaw with traditional design mood boards is that they exist in a vacuum. To fix my corner, ChatGPT needed to see the actual constraints of the room — the lighting, the baseboards, and the proximity to the fireplace.

The process is incredibly seamless. I took a quick, unedited photo of the empty corner on my phone. Opening the ChatGPT app, I tapped the plus (+) icon right next to the chat bar, hit Photos, and uploaded the snapshot.

But a photo is only half the battle; the real magic happens in how you prime the AI. I didn’t just say "make this look nice." I gave it my exact vision. My prompt read:

"I’m redecorating my living room. I want to do something in that left corner to replace the unit, I’m thinking of a comfy chair with a little table, a bookcase or corner unit for books and a standing lamp. Can you show me what that could look like?"

I hit the arrow button to generate, and within about 45 seconds, the AI did what days of Pinterest scrolling couldn’t. It analyzed the dimensions of my photo and served up a fully rendered, highly realistic visualization of my corner. I had some things to tweak, but it was an amazing start.

Refining the design

Seeing the initial layout was a massive relief, but design is always iterative. The AI first served up a great blueprint: a cream armchair on a circular wicker rug next to a black corner unit. The layout was perfect, but the colors felt off for my space.

Before tweaking it, though, I needed to clear a mental hurdle. Every guest to my new home asked if I was putting a dining table there. I didn't want one, but for peace of mind, I had ChatGPT swap the nook for a table and chairs. Seeing it look instantly cramped was the exact validation I needed to kill the idea for good.

With the reading nook confirmed as the winner, I went back to the original image to use ChatGPT’s secret weapon: the live image editing canvas.

By tapping the Edit button on the generated image, I opened a suite of editing tools. Instead of starting from scratch, I typed: "Can you make the rug blush pink, add some more plants to the bookcase, change the bookshelf to a light-colored wood, and try a more speakeasy vibe for the chair?"

The new version it produced swapped the wicker for a soft blush pink, added greenery, lightened the heavy black wood, and changed the chair to a plus, mid-century forest green armchair —allowing me to test risky decor choices before committing.

By treating the interface like a living canvas — resizing elements that felt too crowded, removing decor pieces that felt distracting, and swapping out colors on a whim, I made all of my expensive, risky design mistakes digitally.

When I finally closed the app, the design paralysis was entirely gone. Instead of wasting weekends wandering through furniture showrooms or guessing if a piece would actually fit the space, I had an actionable, visual blueprint. Plus, I'd saved hundreds in in potential returns and mistaken purchases.

I knew exactly what sizes to measure for, what colors to buy, and precisely how to kit out my new space with total confidence.

More from Tom's Guide

• Most people miss these 3 ChatGPT tricks — and they dramatically improve your results
• How to get started using ChatGPT: Everything you need to know to make the chatbot work for you
• 7 AI mistakes I wish someone had told me about when I started

Apple on Friday filed a lawsuit against OpenAI, alleging that the artificial intelligence company engaged in the theft of trade secrets, intellectual property and actively coached departing Apple employees how to evade the company's security policies.

"At every level, from members of its Technical Staff to its Chief Hardware Officer, and in coordination with business partners, OpenAI has been stealing Apple’s trade secrets and confidential information," the suit says.

The suit, filed in federal court in Northern California, calls for an injunction against OpenAI, as well as unspecified damages resulting from the theft of Apple's trade secrets.

We've gone through the entire lawsuit; here are four of the biggest allegations Apple claims.

Holding on to company property

One of the plaintiffs named in the lawsuit is Chang Liu, who was a Senior System Electrical Engineer at Apple for eight years, and joined OpenAI in January 2026. The suit alleges that Liu realized he was able to still access Apple's confidential files after leaving, and then also instructed another colleague on how she could also pass along information.

"Within hours of leaving Apple, Mr. Liu shared with Ms. Peng, “I still have another computer” he was planning to use to access Apple’s confidential information so he could discuss it with Ms. Peng that evening. Within weeks of departing Apple, Mr. Liu also used Ms. Peng’s Apple-issued work computer that was authenticated to Apple’s network — accessing it while she was still employed at Apple and he was not."

Mr. Liu discovered that, surprisingly, he still could access the Apple’s network repository after leaving Apple, the result of a then-unknown authentication vulnerability. Rather than bringing this to Apple’s attention or protecting Apple’s confidential information as required by his IPA, Mr. Liu celebrated his find with Ms. Peng and set about exploiting it: “LOL, I found out I can access the [network storage], so funny.” Ms. Peng’s response was immediate: “I’m ready.”

Show-and-tell

Apple alleges that, as part of OpenAI's interview process, Tang Yew Tan, who was the Vice President of Product Design for iPhone and Apple Watch, and now OpenAI’s Chief Hardware Officer, asked Apple employees to bring detailed CAD drawings, prototypes and other samples.

"Mr. Tan has directed candidates to bring physical hardware components, “parts,” and product samples from their Apple work to their OpenAI interviews for “show and tell” sessions that would disclose Apple’s proprietary technologies. For example, messages left on an Apple-issued work device show that Mr. Tan instructed an Apple employee to “bring some parts [she] worked on” such as “Batteries,” “SIP” (Systems-in-Package), “mlb” (multi-layer or main logic boards), and “shields” and that it may “be good to show” other interviewers these Apple components."

Going after Apple's vendors

The lawsuit also alleges that OpenAI went after Apple's suppliers to gain additional information about Apple's manufacturing processes and components.

"OpenAI also has been targeting Apple’s prized partner network and supply chain directly. It has been using misappropriated knowledge of Apple’s confidential relationships, manufacturing and design processes, and proprietary terminology to extract from trusted third parties additional trade secret information and processes."

Coaching departing Apple employees

It's no secret that a number of Apple employees have left to go to OpenAI; the suit alleges that the company coached them how to announce their departures in such a way that they could continue to remove confidential information.

"OpenAI has actively coached Apple employees on how to manage their exits from Apple in ways that avoid scrutiny. This helps OpenAI’s new hires evade Apple’s security procedures and reduce the chance that their confidentiality violations and trade secret theft would be detected."

"OpenAI has counseled departing employees not to disclose their next employer and given advice on how to avoid the “dreaded walk out” that would promptly remove them from the company rather than giving them a standard two weeks in which they could continue to access Apple’s confidential information and trade secrets."

The takeaway

Obviously, this is a huge lawsuit with massive implications. As recently as two years ago, the companies were in a partnership to use ChatGPT in Apple products, but as OpenAI has announced its ambitions to get into hardware, there's been increasing friction between the companies — not helped by the fact that a number of Apple employees have gone to work at OpenAI.

Most likely, this suit will take years to settle, but it could drastically shape the fortunes of both companies.

More from Tom's Guide

• Claude isn't just a chatbot anymore — here are 9 things it can do in 2026
• OpenAI is shutting down its AI browser — but ChatGPT users are getting something better
• The end of forgetting: Why AI’s perfect memory is a psychological trap

OpenAI is saying goodbye to Atlas, the AI-powered web browser it launched less than a year ago. But don't mistake the move for a retreat from the web. Instead of trying to convince millions of people to abandon Chrome or Safari, OpenAI is folding Atlas' most useful AI features directly into ChatGPT, its desktop app and a new Chrome extension.

The decision marks a significant shift in strategy, suggesting the company no longer sees AI browsing as something that needs its own browser. Instead, it wants ChatGPT to become an AI assistant that works wherever you already browse.

What happened?

OpenAI confirmed that Atlas will begin shutting down on Aug. 9.

Rather than continue developing a standalone browser, the company says many of Atlas' AI-powered browsing tools, including agentic features capable of interacting with websites on your behalf, will continue inside ChatGPT Work, the ChatGPT desktop app and an upcoming Chrome extension. Existing Atlas users will be guided through the transition over the coming weeks.

While the Atlas name is going away, the technology behind it isn't. Instead, OpenAI is moving those capabilities into products that already have a much larger user base.

Why OpenAI changed course

When Atlas launched, the idea was pretty straightforward: create an AI-first browser that could summarize web pages, answer questions about what you were reading and complete tasks online without requiring constant manual input.

The challenge wasn't the technology but changing people's habits. As a Chrome user, I can honestly say I found it hard to leave behind. And I'm not alone. Chrome remains the dominant browser, while Safari is deeply integrated into Apple's ecosystem. Convincing users to switch browsers is a much bigger hurdle than adding new AI capabilities to software they already use every day.

OpenAI now appears to be betting that AI doesn't need to replace your browser. It simply needs to make the browser you already use smarter.

What this means for ChatGPT users

For most people, this should make AI browsing feel much more seamless. Instead of opening a dedicated browser to access AI features, users will increasingly find those capabilities built directly into ChatGPT or available through Chrome. That means fewer apps to install and manage while keeping the ability to summarize long articles, research complex topics, automate repetitive online tasks and interact with websites using AI.

This also marks a shift across the industry. OpenAI's decision also reflects a broader trend taking shape across the AI industry. Earlier this year, AI-powered browsers were positioned as the next major computing platform. Companies raced to build browsers with built-in assistants, hoping users would abandon traditional web browsers altogether.

That vision now appears to be evolving. Instead of competing to become your primary browser, AI companies are increasingly embedding their assistants into the software people already rely on. Whether you're using a browser, a desktop application or another productivity tool, the goal is the same: make AI available wherever work happens.

Rather than creating yet another destination, AI is becoming part of the tools we already use.

The bottom line

Atlas may be disappearing, but OpenAI's ambitions for AI-powered browsing are only growing.

By moving Atlas' technology into ChatGPT and Chrome instead of maintaining a standalone browser, OpenAI is betting that convenience matters more than another browser icon on your desktop.

If that strategy succeeds, the future of AI browsing won't be about choosing a different browser at all. It'll be about having an intelligent assistant available no matter which browser you prefer.

More from Tom's Guide

• Gemini isn't just a chatbot anymore — here are 13 things it can do in 2026
• I replaced my true crime podcast with ChatGPT Voice for a week — here's what happened
• ChatGPT isn't just a chatbot anymore — here are 11 things it can do in 2026

OpenAI just dropped GPT-Live (specifically GPT-Live-1 and GPT-Live-1 mini), and it completely changes how it feels to talk to an AI. We've all seen what ChatGPT Voice can do, but this feels like a completely different AI.

Previous versions of ChatGPT Voice felt rigid and robotic because it didn't quite understand pauses or when exactly to respond. That's why I can't emphasize enough how much different GPT-Live feels. It completely breaks that mold. By far the best aspect of this update is how it handles continuous, real-time media.

Intrigued by the update's new "full-duplex" tech, I decided to test its limits during the 2026 World Cup knockout rounds. Although there were no games today, I put my phone next to my laptop speakers, pulled up a previous Telemundo stream, and let ChatGPT listen to the rapid-fire Spanish commentary.

Instead of waiting for pauses or cutting out when the announcers yelled, ChatGPT acted exactly like a simultaneous interpreter by providing a continuous, running English translation directly over the top of the broadcast's audio.

Here is how the underlying technology makes this possible, how OpenAI is keeping it safe, and the exact prompts you can use to try it yourself.

What Makes GPT-Live different?

For starters, the biggest technical breakthrough here is a full-duplex architecture. Instead of taking turns, GPT-Live listens and speaks at the exact same time. Not always, as in, it's not talking in English alongside of you. But it mimics human conversation dynamics by dropping active listening cues like "mhmm," "yeah," or "got it" while you talk or simultaneously translates into English when asked. And, if you trail off to think, it actually waits for you instead of cutting you off.

OpenAI calls this decoupled delegation, meaning ChatGPT stays on the line, chatting with you naturally, while delegating heavy-duty reasoning or deep web searches to OpenAI's flagship models (like GPT-5.5) in the background.

Plus, you no longer have to listen to an AI read out a long list of numbers or complex details. While you are talking, ChatGPT can now dynamically pop up rich visual cards on your screen for things like weather forecasts, stock charts, maps and sports stats. I found the visuals to be extremely helpful, particularly for things like dates of the World Cup.

5 Prompts to Put GPT-Live to the Test

Because GPT-Live thrives on interruptions, multitasking and continuous listening, standard text prompts don’t do it justice. Open the ChatGPT app, tap the waveform icon and try these five hands-free prompts:

1. The international sports hack

Try the prompt: "I am about to watch a live sports broadcast out loud in [Foreign Language, e.g., Spanish/French]. I want you to act as my simultaneous translator. Do not interrupt the audio, do not wait for pauses, and do not add any commentary of your own. Just provide a continuous, running English translation of everything the commentators are saying as you hear it."

Why it works: Because GPT-Live can listen and speak at the same time, it can track continuous external audio. And, unlike Meta Display Glasses that require you to look directly at the person speaking, ChatGPT simply listens. So, all you need to do is turn on an international stream or watch post-match press conferences on your TV, set your phone down near the speaker, and get an instant audio dub.

2. The active brainstormer

Try the prompt: "Hey, let's brainstorm some unique business concepts for a local, sustainable coffee shop. I want you to pitch me an idea, but I’m going to interrupt you with objections as you speak, and I want you to pivot your pitch on the fly based on what I say."

Why it works: This tests the model's full-duplex capabilities. You can cut it off mid-sentence, and it will smoothly adapt its logic without stumbling or restarting its thought process.

3. The multi-tasking match tracker

Try the prompt: "I'm watching the World Cup right now. Can you look up the current tournament bracket standings, find a quick list of upcoming quarterfinal matches, and check the weather forecast for Miami's stadium? Keep talking to me about the tournament's biggest upsets while you look it all up."

Why it works: This forces the model to use decoupled delegation. It will keep chatting with you casually while background processes fetch real-time data, triggering the new visual cards right on your screen.

4. The 'don't interrupt me' brain dump

Try the prompt: "I'm going to explain a complex problem I'm having with a project layout. I tend to hesitate, pause, and talk through my thoughts slowly. I want you to just listen actively, acknowledge that you're following, and do not give me a final answer until I explicitly say 'Okay, I'm done, what do you think?'"

Why it works: This tests the model's new turn-detection and background noise filtering. It proves it can tolerate natural human pauses without jumping the gun.

5. The live language coach

Try the prompt: "Act as my live translator. I'm going to practice ordering food at a restaurant in Italian. Let's do a fast-paced roleplay where you play the waiter, speak completely in Italian, but give me quick, subtle English audio hints if I hesitate for too long on a word."

Why it works: The ultra-low latency and fluid back-and-forth make real-time language practice feel incredibly organic rather than robotic.

Safety designed for real-time speech

Because audio-native conversations carry more emotional weight and happen faster than text, OpenAI introduced entirely new safeguards alongside GPT-Live:

• Real-time 'runtime' controls: The system monitors the conversation as it happens. If the AI veers into a risky generation, it will dynamically steer the voice response away, or terminate the call entirely in high-risk cases.
• Teen protections & parental controls: OpenAI says age-appropriate behavior is trained directly into the model. Parents can completely disable ChatGPT Voice for their teens via Parental Controls, and linked parents can be notified if a teen exhibits signs of severe distress or self-harm.
• Anti-impersonation: To prevent deepfakes, GPT-Live is strictly locked to OpenAI’s nine remastered, predefined system voices. OpenAI says it will reject any command to clone or imitate a real person’s voice.
• Monitoring emotional reliance: Because the model sounds incredibly human (using conversational filler like "mhmm"), OpenAI has deployed long-term tracking to study and prevent users from forming unhealthy emotional attachments to the AI.

How to get it

GPT-Live-1 is rolling out globally now as the default voice engine for ChatGPT Plus, Pro, and Go users, while GPT-Live-1 mini is hitting the Free tier. If you have the update on iOS, Android or desktop, just tap the voice button and experience the next generation of AI interaction for yourself.

You'll notice the typical blue circle is now lavender, which is how you know you're using GPT-Live-1. Let me know what you think in the comments.

More from Tom's Guide

• Gemini isn't just a chatbot anymore — here are 13 things it can do in 2026
• I replaced my true crime podcast with ChatGPT Voice for a week — here's what happened
• ChatGPT isn't just a chatbot anymore — here are 11 things it can do in 2026

More from Tom's Guide

• Gemini isn't just a chatbot anymore — here are 13 things it can do in 2026
• I replaced my true crime podcast with ChatGPT Voice for a week — here's what happened
• ChatGPT isn't just a chatbot anymore — here are 11 things it can do in 2026

More from Tom's Guide

• Gemini isn't just a chatbot anymore — here are 13 things it can do in 2026
• I replaced my true crime podcast with ChatGPT Voice for a week — here's what happened
• ChatGPT isn't just a chatbot anymore — here are 11 things it can do in 2026

We've come a long way since ChatGPT first launched. From smarter and faster models to an overall more natural conversation, it's hard to believe the model was just launched a few years ago.

But approaches that made sense in 2022 simply don't work as well now. Today's ChatGPT is much better at maintaining context, reasoning through complex problems and collaborating over multiple back-and-forth messages. Yet many people are still using it the way they did years ago, missing out on some of its biggest improvements.

If your ChatGPT conversations haven't evolved, these are the five habits I'd leave behind.

1. Stop trying to write one perfect prompt

One of the biggest myths that refuses to die is that every ChatGPT conversation has to begin with a massive prompt containing every possible detail.

It doesn't, especially now that ChatGPT is better at understanding context. In fact, I've found I get better results by treating ChatGPT like a collaborator instead of a vending machine that spits out responses.

For the best results, start with a simple request, look at what it produces and then refine it together over a few follow-up messages. When you do this, the conversation actually becomes part of the creative process, and the final result is usually much stronger than trying to predict every requirement upfront.

2. Stop treating ChatGPT like a search engine

If you're only asking ChatGPT for quick facts with Google-style searches, you're barely scratching the surface.

Search engines are great for finding information and considering how much energy AI uses, it's good to keep using Google or favorite search engine for that. But ChatGPT shines when you need help thinking through information.

That means, instead of asking, "What's the best laptop under $1,000?"

Try asking: "Compare these three laptops under $1,000 for a journalism student who travels frequently. Explain the tradeoffs, recommend one and tell me what I'd be giving up by choosing it."

Or instead of requesting a definition, ask ChatGPT to compare ideas, explain why experts disagree or summarize multiple viewpoints before making a recommendation. The more you ask it to analyze and synthesize, the more valuable it becomes.

3. Don't ignore Memory if it fits your workflow

One feature I think many people overlook is Memory. When enabled, ChatGPT can remember details like your writing style, recurring projects or personal preferences, so you don't have to repeat the same instructions every time you start a conversation.

For example, if you regularly ask ChatGPT to write in a conversational style or help with the same ongoing project, Memory can save a surprising amount of time.

Of course, it's not for everyone. If you'd rather every conversation start fresh, you can turn Memory off or manage exactly what ChatGPT remembers. The important thing is knowing the feature exists and deciding whether it makes sense for how you use ChatGPT.

4. Don't settle for the first answer

One of the best prompts I use comes after ChatGPT has already answered. Instead of copying the first response, I'll ask things like: "Critique your own answer." Or, "What's missing?"

In most cases, the first response is often just a starting point. Some of my favorite ChatGPT conversations have gone through five or six rounds of refinement before I was happy with the result.

5. Stop using the same ChatGPT tool for every task

Many people still think of ChatGPT as one AI assistant. In reality, it's become a collection of specialized tools.

If I'm researching a complicated topic, I'll use Deep Research. If I'm solving a difficult reasoning problem, I'll choose a reasoning-focused model. If I need an illustration or concept image, I'll switch to image generation. For long-term projects, I'll organize everything inside Projects instead of starting over from scratch every time.

Choosing the right capability for the job often makes a bigger difference than rewriting your prompt another dozen times.

ChatGPT has changed, so should your approach

The biggest change is that ChatGPT has become much more of a collaborator. But for that reason, the habits that worked during ChatGPT's first year aren't necessarily the ones that produce the best results today. Instead of chasing the perfect prompt, think of ChatGPT as a partner you can brainstorm with, challenge, refine and revisit over time.

Once you make that shift, you'll probably find yourself spending less time fighting the AI and more time getting exactly what you need.

More from Tom's Guide

• Gemini isn't just a chatbot anymore — here are 13 things it can do in 2026
• I replaced my true crime podcast with ChatGPT Voice for a week — here's what happened
• ChatGPT isn't just a chatbot anymore — here are 11 things it can do in 2026

Most of today's AI assistants can use at least some customer conversations to improve future models, although the rules vary dramatically depending on which service you're using, whether you have a personal or business account, and which privacy settings you've enabled.

After reading about Google's latest changes around Search and AI training, I decided to spend some time checking the privacy controls for every major AI chatbot I use regularly.

I wanted to know exactly which companies actually make it easy to control how your data is used. I did the work so you don't have to — here's how they actually compare.

Not every AI company plays by the same rules

It's tempting to believe that all chatbots have similar privacy rules considering how similar they all seem. But behind the scenes, they're collecting and handling your data in very different ways.

Some make it easy to opt out of having conversations used to improve AI models. Others bury the controls several menus deep. And in some cases, there isn't a meaningful opt-out at all.

After comparing ChatGPT, Gemini, Claude, Microsoft Copilot, Meta AI and Grok, one thing became obvious: there is no industry standard for AI privacy.

ChatGPT: The most straightforward

The biggest surprise for me was ChatGPT. I was not expecting the settings to be as easy to understand. Quite the contrary, I expected the settings to be buried somewhere inside a privacy dashboard. Instead, they're right where most people would think to look.

Open Settings > Data Controls and you'll see a switch labeled "Improve the model for everyone."

Turn it off, and OpenAI says your future conversations won't be used to improve its models.

You can also use Temporary Chat, which doesn't save conversations to your history or use them for model training.

It took less than a minute to change. Whether you're comfortable sharing conversations or not, I appreciated how straightforward the process was.

Google Gemini: Transparent, but with a tradeoff

Google's controls weren't especially difficult to find either. Gemini Apps Activity lets you decide whether your conversations are stored and used according to Google's policies.

The catch is that disabling activity also changes how Gemini works. You'll lose saved chat history, and some personalized experiences disappear along with it. That's not necessarily a bad tradeoff, just something you'll have to weigh before making the switch.

As mentioned, Google recently confirmed that some Search interactions — including uploads used with AI-powered Search features, may contribute to improving its AI systems unless you disable the relevant activity settings.

Claude: Clear controls, with some fine print worth reading

Anthropic introduced a training opt-in for its consumer Claude products (Free, Pro and Max) in August 2025. Users were asked to decide whether their conversations could be used to improve future models, with a deadline to make their choice.

The setting lives under Settings > Privacy, where you'll find a toggle labeled "Improve Claude for everyone." Turn it off and Anthropic says your future conversations won't be used for model training.

Like ChatGPT, the change is forward-looking — it doesn't pull your data out of models that have already been trained. And if you leave the toggle on, Anthropic's retention period for that data extends from 30 days to up to five years.

There's also an Incognito mode, accessible via a ghost icon in the interface. Incognito conversations aren't used for training regardless of your main toggle setting, which is a useful option for one-off sensitive queries.

One thing worth noting: Anthropic's privacy policy includes a carve-out for conversations flagged by its safety classifiers. Even if you've opted out, flagged content may still be used to improve Anthropic's internal trust and safety systems. The policy doesn't define exactly what triggers a flag or commit to notifying users when one occurs.

Business and API users operate under separate commercial terms, and their data isn't used for model training by default.

Microsoft Copilot: It depends on which product you're using

Microsoft was the hardest to summarize because the answer changes depending on where you're using AI.

For the consumer Copilot experience (used with a personal Microsoft account), Microsoft does use conversation data — including voice interactions and uploaded files' context — to train its AI models by default.

However, an opt-out exists. On copilot.com or the desktop app, go to your profile icon > your profile name > Privacy >Training on conversation activity.

On mobile, the path is Menu > profile icon > Account > Privacy > Training on conversation activity. Opting out excludes your future conversations from model training.

One detail that separates Microsoft from the pack is that even if you don't opt out, Microsoft says personal identifiers are anonymized before data is used for training. Uploaded files themselves are never used for training and are deleted after 30 days.

For Microsoft 365 Copilot used through a work or school account, the picture is much cleaner. Microsoft explicitly states that enterprise data is not used to train its foundation AI models. Your organization's IT administrator controls the relevant settings, and users in those environments won't even see a training toggle.

The gap between these two experiences is significant because a personal user who doesn't dig into settings is opted in by default. An enterprise user is protected by contract. If you're using Copilot casually with a personal account, it's worth the two minutes to check.

Meta AI: No simple opt-out for most users

Meta AI is where the comparison gets uncomfortable. Unlike every other service on this list, Meta doesn't offer U.S. users a simple toggle to opt out of AI training. Your public posts, comments, captions and interactions with Meta AI across Facebook, Instagram and WhatsApp can all be used to train Meta's generative AI models.

EU and UK users have stronger protections, they can submit a formal objection through Meta's Privacy Center under GDPR rights, and Meta is required to review it.

But for users in the United States, the only available option is a third-party information form designed for cases where your personal data appeared in an AI-generated response. Meta doesn't guarantee it will honor these requests.

There's also no way to submit an opt-out request specifically for Meta AI interactions on WhatsApp. Meta says private messages aren't used for training, but any conversation you have directly with Meta AI can be.

Making your accounts private reduces your exposure, since Meta says it doesn't scrape private posts. But that's a blunt instrument, it changes your entire social media experience to address an AI training concern.

To make matters more complex, as of late 2025 Meta began using AI chat interactions for ad targeting across its platforms, with no opt-out available for that specific use.

Grok: An opt-out exists, but the scope is wider than you'd expect

Grok, built by xAI and integrated into X (formerly Twitter), does offer a training opt-out, but the situation is more complicated than it first appears because there are actually two separate data flows to manage.

First, there's your Grok conversation data. You can disable training use by going to Settings > Data Controls within Grok and turning off the relevant toggle. xAI has confirmed that once disabled, your future chats won't be used for model training.

Second, there's your X platform data. By default, your public posts, likes and interactions on X are also used to train Grok, regardless if you've ever opened the chatbot. To disable this, go to Settings and Privacy > Privacy and Safety on X and find the Grok data-sharing section.

Both settings need to be turned off separately. Disabling one doesn't affect the other.

That dual structure is the key difference between Grok and standalone chatbots like ChatGPT or Claude. Those services only train on what you tell them directly. Grok can train on everything you've ever posted publicly on X, and opting out doesn't retroactively remove data that's already been collected.

Grok also offers a Private Chat mode, where conversations aren't saved and are deleted within about 30 days, similar to incognito features on other platforms. But the broader X data pipeline makes the overall privacy picture murkier than the conversation-level toggle alone would suggest.

What I learned

At one end, ChatGPT and Claude offer clear, easy-to-find toggles that take under a minute to change. Google is similarly accessible but comes with a functional tradeoff. Microsoft splits the difference depending on whether you're a personal or enterprise user.

At the other end, Meta offers U.S. users no meaningful conversation-level opt-out, and Grok requires managing two separate settings across two different interfaces to fully limit your exposure.

As AI becomes part of everyday life, it's important to really understand the privacy settings. Choosing a chatbot may involve more than deciding which one writes the best email or creates the best image and come down to a much simpler question: Which company do you trust with your conversations and how hard do they make you work to answer that for yourself?

Let me know in the comments which chatbot you prefer and why.

More from Tom's Guide

• Gemini isn't just a chatbot anymore — here are 13 things it can do in 2026
• I replaced my true crime podcast with ChatGPT Voice for a week — here's what happened
• ChatGPT isn't just a chatbot anymore — here are 11 things it can do in 2026

ChatGPT has come a long way since it launched. Even within the last year, we've seen significant improvements in Memory, speed and even more app integrations. Simply put, the platform has become an incredible daily workspace. I get so much done within the space because of its built-in functionalities.

The AI can organize long-term projects, run heavy research, connect directly to my other apps, handle coding tasks, remember my preferences and it can even break down my financial planning.

So when friends ask me the best prompt for writing emails, I respond with a good prompt but then tell them they're completely missing the tools that actually change how daily work gets done. Here are seven features worth exploring right now, along with exactly how the pricing works for each.

1. Projects keep my workspace organized

I am the type of person that takes hand-written notes in meetings and style writes reminders on sticky notes. The brain-hand combination works for me. At the same time, I need those notes to not get lost on my desk. For that reason, and a plethora of others, I've found Projects to be a game-changer for my workflow.

Whether I'm uploading messy notes or keeping track of related chats, Projects helps to maintain context. I upload documents, add custom instructions and keep everything for a specific goal in one centralized place.

I also rely on Projects to manage long-running work. I keep separate spaces for planning vacations, organizing home renovations, managing my writing and running work on side projects. ChatGPT already understands the background context every time I log in.

Pricing: Projects is a core feature, though the file upload limits scale with your specific plan.

2. Deep Research builds full reports

Sometimes I want to do more than simply skim the internet. It's times like those when I don't need an answer in seconds. Deep Research spends significantly longer gathering information, comparing sources, and building a comprehensive report.

ChatGPT does the heavy lifting of opening dozens of tabs and presents its findings with source citations for verification. I've used it to find the best value electric SUVs, compare mortgage options, analyze competing software and understand my own family's history.

When Deep Research first rolled out, I didn't think I'd ever have a use for it. Now, I use it nearly every day.

Pricing: Free users get 5 lightweight queries per month. Plus users receive 10 full queries, the $100 Pro tier gets 50, and the $200 Pro tier has unlimited access.

3. Apps connect directly to my workflow

OpenAI recently streamlined third-party integrations into an "Apps" directory. By connecting supported cloud services and workspaces, ChatGPT works across the tools I already use.

I ask a single question, and the platform finds specific files in my cloud storage, searches email threads, reviews documents, and references my calendar for scheduling. It saves me hours of bouncing between different applications.

Pricing: App connections are available across plans, but advanced integrations and higher usage limits require a Plus, Team, or Enterprise subscription.

4. The Codex app automates my desktop

Codex has grown from an underlying model into a dedicated desktop agent workspace. It's like a collaborator that can build software, fix or explain code and automate repetitive tasks directly on your machine.

I use it as a massive productivity boost. Even for someone who has never written a line of code, Codex bridges the gap between an idea and a working project. It helps me create simple websites, small desktop utilities, spreadsheet macros and custom workflow shortcuts.

Pricing: Codex is generally available on all plans. Paid subscriptions offer more daily agent-turn hours and higher rate limits for complex automations.

5. Voice conversations help me brainstorm

I use Voice ChatGPT more than I'd like to admit. It's a way to brainstorm ideas with AI when I'm working on a quick idea. I don't have to slow down to type, I simply just use Voice Mode to challenge my ideas, consider new ones rather than try to organize messy thoughts on the screen.

Speaking naturally feels like collaborating with another person.

I use voice mode to brainstorm article ideas out loud, practice rapid-fire interview questions, talk through difficult decisions and learn unfamiliar topics via a two-way dialogue.

Pricing: Standard voice features are free. The highly conversational "Advanced Voice Mode" is strictly time-capped for free users, making a Plus or Pro subscription necessary for long conversations.

6. Memory tailors my results

Memory is a massive quality-of-life improvement. ChatGPT remembers information I choose to share, including my preferred writing style, recurring projects, and dietary restrictions.

I spend less time repeating myself and more time getting tailored answers. I remain fully in control and can review, edit, or delete saved memories at any time. Personalization remains a choice.

Pricing: Memory is available on the Free tier. Plus and Pro users benefit from higher overall message limits, allowing the AI to process and retain complex context more reliably.

7. Financial explanations clarify my budget

I'll be honest, I was hesitant to try this feature. Handing over my financial activity to AI seemed like a huge mistake. Luckily, I explored this feature a little more and discovered I could use it without handing over all my private infromation. Instead of using ChatGPT to make ative investment decisions, I use it exclusively as a financial educator.

I ask the platform to explain unfamiliar investing terms in plain English, compare budgeting methods, translate dense insurance jargon and outline the differences between various retirement accounts. It helps me build realistic savings plans based on hypothetical budgets.

AI shouldn't replace a human finacial advisor and you shouldn't give your personal information, but it can be helpful otherwise.

Pricing: General financial questions are free. Uploading large budgeting spreadsheets for deep data analysis quickly hits the Free tier limits, making a Plus subscription ideal for heavy number-crunching.

ChatGPT has become indispensable

People often ask me what has changed the most about ChatGPT. While it's hard to pinpoint just one thing, my response is always that the AI has become far more useful.

Today, ChatGPT organizes long-term work, conducts deep research, connects with external apps, automates tasks through Codex, holds natural voice conversations and explains complicated topics.

Give these uses a try and let me know what you think in the comments. I'd enjoy knowing how you use AI.

More from Tom's Guide

• Gemini isn't just a chatbot anymore — here are 13 things it can do in 2026
• I replaced my true crime podcast with ChatGPT Voice for a week — here's what happened
• ChatGPT isn't just a chatbot anymore — here are 11 things it can do in 2026

The future of Apple's rumored smart glasses and the existing Vision Pro headset has become blurrier in recent weeks. Between price hikes, the alleged back-burning of projects and high-level departures, it seems from the outside that Apple's Vision division isn't working.

The most recent potential setback came last week when Bloomberg's Mark Gurman reported that Paul Meade, the vice president in charge of the Vision Pro and glasses efforts, is leaving for OpenAI. Meade is the second Vision executive to leave in the last year after the creator of the Vision Pro, Mike Rockwell, was shifted to overseeing Siri after AI failures in that department.

Meade, meanwhile, was an integral part of Vision Pro's development and has reportedly been responsible for developing the Apple Glasses, as per Gurman's sources.

Normal attrition or an indicator of problems?

Apple is in a period of transition as Tim Cook's reign as CEO comes to an end and John Ternus starts his takeover. Soon after Ternus was announced as the heir apparent, Johny Srouji was promoted to chief hardware officer, and quickly went about making controversial changes to Apple's product design org chart.

Analyst Avi Greengart of Techsponetial told Tom's Guide that this kind of attrition is normal, especially after a leadership change: "The executive moves that Apple is making appear to be more about securing the future of Siri – which impacts nearly every product Apple makes."

As MacRumors reports, several members of Apple's leadership team have been there for decades, including marketing head Greg Joswiak and services chief Eddy Cue. In the coming years, we may see more departures as the old guard retires or moves on and younger faces rise up.

The question is whether the current attrition is a new development strategy on Ternus' part, which opened the door for people like Meade to leave. Or is it a more challenging problem?

"Sometimes that's the way companies move forward," IDC analyst Francisco Jeronimo told Tom's Guide. "Potentially, [Apple] is rethinking its future product roadmap and looking for a different challenge."

Jeronimo noted that a problem that Apple may have is a Jony Ive one.

OpenAI and Jony Ive poaching

As Paul Meade leaves for OpenAI it looks like former Apple designer Jony Ive is rebuilding his Cupertino team at his new home. Meade joins Tang Tan and Evans Harkey, who used to run design, hardware, product design and industrial design at Apple with Ive. Gurman noted that Ive has also pillaged Apple's rank-and-file designers and engineers to build out his team.

That trio brings a lot of expertise and knowledge of Apple's plans to OpenAI, which is in the midst of designing its own suite of AI-forward devices.

The most recent rumors indicate that OpenAI has "fast-tracked" development of its own ChatGPT phone, which could debut as soon as the first quarter of 2027. Reportedly, OpenAI and Ive are developing at least five other AI-powered products, including earbuds, speakers, smart glasses, a smart pen, and a Humane AI-esque pin.

OpenAI promised AI devices that will replace the phone, though IDC's Jeronimo said that any other kind of device will be harder for consumers to think they need. "There are concerns that it will not going to be as popular as they [OpenAI] think."

Where does this leave Apple Glasses?

Apple has reportedly back-burned its expensive, heavy Vision Pro headset in favor of two different smart glasses. Supply chain analysts have claimed the Meta Ray-Ban-esque Apple Glasses will launch in 2027, while an XR version might debut in 2028 or 2029.

"I expect Apple to be a player in AI wearables one way or another," Greengart told Tom's Guide.

Everyone I spoke with agreed that Apple Glasses seem to be the next step; it's just unclear what direction Apple wants to take them. Though any glasses will likely be more consumer-friendly and cheaper than the Vision Pro.

If Apple gets glasses correct, it could be the next Apple Watch and actually send shockwaves through the eyewear industry. The question is whether key departures are draining Apple of talent and knowledge or just an adjustment of priorities.

More from Tom's Guide

• 'Easily the biggest leak in Apple's history': iPhone 18 Pro final design may have just been revealed in a stolen drop test video
• Apple seeks RAM from blacklisted Chinese company as memory shortage impacts everything — but it may not be enough
• Apple MacBook Ultra may arrive running on some of the same chips as MacBook Pro — here's why

A new leak reported by Crypto Briefing suggests OpenAI is working on a bidirectional voice model that could make conversations with ChatGPT feel much more natural.

Instead of waiting for you to finish speaking before responding, the AI would reportedly be able to listen while it's talking — allowing for interruptions and a flow that mimics real conversation between two people.

On paper, that sounds like a nice upgrade, but I think highlights a growing trend towards something much bigger.

AI speaking is upon us

For the past few years, we've been talking about how AI is changing the way we write. I've felt it in my own work. I studied writing long before AI existed and today it feels like I've had to learn the craft all over again. My sentences are shorter. I cut clever turns of phrase to literally avoid sounding like AI. I second-guess punctuation I used to love. Not because my taste changed, but because writing for an internet increasingly filtered through AI rewards a different style.

I've accepted that. What worries me now isn't just AI changing how we write, but how we speak.

If leaks like this become reality — and I suspect they will — we're heading toward a future where interacting with AI won't require a keyboard at all. All we'll have to do is talk. Whether we're chatting through a pair of the best smart glasses or with ChatGPT in our cars, the screen will soon become optional.

And while that sounds incredibly convenient. It also raises a question I can't stop thinking about: if we stop typing, do we eventually stop reading?

What happens when info is heard instead of read

When information is spoken instead of written, when AI summarizes every article, explains every concept and answers every question conversationally, do we lose the habit of wrestling with difficult text ourselves?

I don't think books disappear. But I do wonder if reading becomes more like handwriting or cursive, which is a valuable skill that fewer people practice because technology offers an easier alternative. And while calculators didn't eliminate math, and GPS didn't make maps disappear, they did change what people practiced every day, and the skills that atrophied were the ones technology made optional.

Reading is different because it isn't just about absorbing information. It's how we build vocabulary, strengthen attention spans, develop empathy and learn to think through complex ideas over long stretches of time. Those are cognitive muscles that grow through practice, and they weaken without it.

Ironically, AI may also be one of the best tools we've ever had for preserving those skills. Imagine an assistant that helps you work through a difficult chapter, quizzes you on what you've read or recommends books based on your favorite authors.

What makes this leak different is that it isn't happening in isolation

What makes this leak different is that it isn't happening in isolation. Over the past year, nearly every major AI company has been racing toward a screenless AI.

OpenAI has made ChatGPT Voice feel increasingly conversational while Google is weaving Gemini into Android, earbuds and smart glasses. Meta is betting heavily on AI-powered Ray-Ban glasses that can see what you see and answer questions in real time. Apple is slowly repositioning Siri as a more capable AI assistant that lives across your devices.

What I'm seeing is that we are moving away from an on-screen chatbot to a world where AI becomes almost ambeint. If AI can listen while it's speaking, understand interruptions and respond naturally without awkward pauses, it removes one of the biggest barriers between humans and machines. The conversation starts to feel... normal.

Final thoughts

Once talking becomes easier than typing, most people will choose talking. And that's the future I keep coming back to when I think about AI. While the technology is remarkable, it's different and ultimately changing how we interact with AI — even more human-like.

And while every technological shift changes not only what we can do, but what we choose to practice, if voice becomes our primary interface with AI, then reading and writing may quietly become secondary skills for many people.

That possibility fascinates me as much as it worries me.

More from Tom's Guide

• I almost ruined a relationship with one text — then I tried this ChatGPT prompt
• I built five ChatGPT agents — then replaced them with one 'master' AI
• Even Google doesn't have enough AI power — and that should worry all of us