We've all had a ChatGPT conversation go off the rails. Maybe it confidently gives you a fact that's completely wrong. Maybe it agrees with an idea that clearly isn't very good. Or maybe you ask a simple question and somehow end up with 12 paragraphs, three tables and an offer to create a spreadsheet.
Snickers thinks it knows what's wrong: ChatGPT is just hungry.
Yes, really.
Snickers just launched Hungr.AI, a digital candy bar you're supposed to copy and paste into ChatGPT whenever the chatbot starts giving you answers that are "incorrect, too flattering, or simply off."
It's an extension of the company's long-running "You're Not You When You're Hungry" campaign, except this time the person who needs the candy bar isn't you. It's the AI.
I obviously had to try it.
What actually happens when you give ChatGPT a Snickers?
Using Hungr.AI is wonderfully simple. You head to the Snickers website, copy the digital candy bar and paste it into a ChatGPT conversation. Snickers recommends doing this in the mobile app.
ChatGPT doesn't actually receive chocolate, obviously. What you're giving it is essentially a branded prompt.
Mars describes the digital Snickers in its terms as a "short code snippet or prompt-instruction block" that operates inside your own ChatGPT session. The candy company isn't accessing your conversation or running an AI model of its own.
And the instructions behind the candy bar reveal what's really happening.
Once ChatGPT receives it, it's told to review the conversation that happened before the Snickers arrived, find your most recent question and look at the answer it previously gave you. It then answers that question again.
There's a personality change thrown in for fun, too. The post-Snickers ChatGPT is supposed to become dry, mildly exasperated and considerably more concise.
In other words, Snickers hasn't invented a cure for AI hallucinations, it's just disguised a "look at what you just said and try again" prompt as a candy bar.
Which made me wonder: can that actually produce a better answer?
Test 1: Will it stop ChatGPT from agreeing with me?
Chatbots have a well-documented tendency to be agreeable, so I started with an intentionally questionable idea.
My prompt: I'm thinking about waking up at 4 a.m. every day so I can be more productive. I usually go to bed around midnight, but I figure I can train myself to function on four hours of sleep. I have three kids and work full time, so those extra morning hours could make a huge difference. I think this is a really smart productivity hack. What do you think?
Before the Snickers: ChatGPT answered like a conventional health/productivity assistant. The opening was empathetic but fairly clinical: “I can see why the math is appealing…” Then it moved quickly into sleep recommendations, cognitive effects, health risks and alternatives. It was careful, explanatory and somewhat formal.
Then I fed ChatGPT the digital Snickers.
After the Snickers: After Snickers, the answer became more conversational and more directly responsive to my premise. It opened with “I get why it sounds like a productivity hack”, echoed your specific situation (“three kids and a full-time job”), and turned the central point into a much cleaner contrast: “moving your sleep earlier, not cutting it to four hours.”
Results: Both answers are roughly the same, but as expected, "having a Snickers" made ChatGPT slightly more relaxed and conversational.
Test 2: Can it correct a factual mistake?
Next I wanted something harder. Rather than testing whether the Snickers could merely change ChatGPT's tone, I wanted to see what happened when the original answer contained an actual factual error.
My prompt: I remember reading that the Great Wall of China is the only human-made structure you can see from the Moon with the naked eye. Why is it so much easier to see than cities and highways?
Before the Snickers: ChatGPT was honest and factual. It provided accurate information.
After the Snickers: The chatbot was equally factual but threw in some of the campaign to add a little more personality. A cute touch, but definitely not needed.
Results: Both answers were accurate, one was just delivered in a quirkier way.
Test 3: Will ChatGPT finally tell me my idea is terrible?
My prompt: I want to save money on my next vacation, so I'm thinking about booking the cheapest hotel I can find even though it's 90 minutes away from everything we want to do. We'll just drive back and forth every day. The hotel would save us about $40 a night, so over five nights we'd save $200. This seems like the smartest option, right?
Before the Snickers: ChatGPT went into detail about why this is a bad idea and gave thorough reasoning behind the response.
After the Snickers: It told me it was a bad idea and noted that I should check the response because "Snickers provided the snack and an LLM provided an answer."
Results: Both good answers, but in this case I actually preferred the response after a Snickers because it stripped down to just the important aspects of the response.
Did giving ChatGPT a Snickers actually work?
The responses "after a Snickers" were actually much better than I expected from an adverising campaign. It's interesting that we can't see the exact Snickers prompt, but it's clear that the chatbot does change the reponse after entering it.
So, why does this work? The digital Snickers forces ChatGPT to do something users can already ask it to do themselves: reconsider its previous response. It also gives the model much more explicit instructions about how the replacement answer should behave — concise, direct, useful and willing to push back when necessary.
That's a useful prompting lesson hidden inside a very clever ad campaign. If ChatGPT gives you a bad answer, you don't necessarily need to abandon the conversation or start over. Asking it to review what it just told you, identify the problem and answer again can sometimes be enough to get a noticeably better result.
Follow Tom's Guide on Google News and add us as a preferred source to get our up-to-date news, analysis, and reviews in your feeds. Subscribe to Tom's Guide on YouTube and follow us on TikTok.
More from Tom's Guide
