I thought AI was getting smarter — until I saw the code that proves it’s ‘razzmatazzing’ us

I use AI every day for work because it helps me brainstorm faster, think clearer and stress-test ideas. At this point, it feels like an assistant that I can lean on at any time.

So when I saw that developer Wes Bos had dug into the source code of Claude and found something unexpected, I assumed it would be a complex breakthrough in neural mapping.

But it wasn’t; what he discovered was something much simpler known as "spinner words." The more I looked into it, the more it changed how I think about the "intelligence" we interact with every day.

The ‘razzmatazzing’ discovery

According to Bos, he uncovered 187 hardcoded verbs tucked away inside an AI system’s code— including words like: “Hullaballooing” and “Razzmatazzing.” Seriously, how great are these words?

But these silly words aren't "hallucinations" or creative bursts from a Large Language Model (LLM). They were pre-written variations — essentially a list of interchangeable words the system pulls from a shelf to avoid sounding repetitive while you wait for a response.

Essentially, some of the most "human" parts of AI aren’t being generated on the fly. They’re chosen from a script.

What are ‘Spinner Words’ anyway?

If you’ve never heard the term, the concept is simple. To keep a user from getting bored or noticing a pattern, a system swaps in different phrases from a predefined list. Instead of the AI always saying "Processing..." it might rotate between: "Hang tight," "Let me think," "Cooking something up" ...and apparently, "Razzmatazzing."

But it's not intelligence, it's purely product design.

And at first glance, "razzmatazzing" feels like a quirky Easter egg. But it points to a reality we often forget: AI isn’t one single "mind"—it’s a layer cake.

When you use a tool like ChatGPT or Claude, you are interacting with:

• The LLM: The actual "brain" that predicts the next word.
• The Guardrails: Safety filters that prevent the AI from going off the rails.
• The UI/UX Wrapper: The "personality" layer where spinner words live.

Spinner words exist to make the AI feel more human-like because they give the illusion of effort and personality while the actual processor is doing the heavy lifting in the background.

Buried in the same discovery was something far more revealing: an analytics system that logs your prompt as "negative" when you swear at it.

The system isn't just generating text; it’s quietly tracking signals like:

• Frustration: Using profanity or caps lock.
• Dissatisfaction: Explicitly telling the AI "This is wrong."
• Failed Responses: When a user has to regenerate an answer multiple times.

When you type something in anger, it’s flagged. Not to "punish" you, but to measure where the product is failing. It’s a reminder that while the AI is helping you work, you are also a data point helping the AI get "better" — or at least, more convincing. That's why when you stop being polite to a chatbot, you might get better answers.

The bottom line

What this discovery means for me, is that I find myself asking was this generated or was this chosen based on how it was prompted? Since seeing this discovery, I have found myself using AI more intentionally.

AI feels human for a reason. It’s not just because the math is getting smarter; it’s because the interface has been carefully designed to sound relatable. Sometimes that means advanced reasoning that changes the world, but sometimes, it just means a hidden list of words meant to keep you smiling while the gears turn.

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.

More from Tom's Guide

• I asked ChatGPT to help me fix 7 everyday problems — I didn’t expect this
• I used ChatGPT wrong for months — these 3 simple 'system' prompts changed everything
• I stopped writing complex AI prompts — this 60-second 'memory' trick works 10x better