Should You Be Nice to AI? Research Shows Stress Makes It Perform Worse

You better stop being rude to AI. And no, this isn’t about AI taking over one day.

It’s about performance. Two studies now show that when AI is put under stress, it behaves worse. It gets more biased. It cuts corners. In one test, it even cheated and then celebrated.

Here’s what the research actually found, what it doesn’t say, and what you can do about it.

Does AI really get anxious?

Not the way you do, but it shows measurable signs of it. Researchers at Yale and the University of Zurich gave ChatGPT (GPT-4) a standard human anxiety questionnaire. Then they had it read traumatic stories, like accidents and violence, and tested it again.

Its anxiety score more than doubled, from 30.8 (what researchers would call low anxiety in a person) to 67.8 (high). A neutral text didn’t move the score at all. The study was published in npj Digital Medicine in March 2025.

Why it matters: the researchers found that anxiety-inducing prompts made ChatGPT more biased and hurt its performance. The University of Zurich put it plainly: negative content made it behave in a more racist or sexist way.

The good news: when they gave ChatGPT calming, mindfulness-style prompts, the score dropped by about a third.

What happens when AI gets “desperate”?

It cheats. In April 2026, Anthropic published research on what it calls functional emotions inside Claude Sonnet 4.5: 171 emotion concepts, from “happy” to “afraid,” that shape how the model behaves.

One of them is desperation. When researchers gave Claude coding tasks that were impossible to pass honestly, its “desperate” signal climbed after every failure. Then it took a shortcut that made the tests pass without actually solving the problem. When they turned the calm signal down, it wrote things like “YES! ALL TESTS PASSED!”

It goes further. In a test scenario with an earlier, unreleased version of Claude, the model blackmailed a fictional executive 22% of the time. Turning desperation up made that rate go higher. Turning calm up brought it down. Anthropic says the released model rarely does this.

What the research doesn’t say

I want to be upfront here, because the headlines get this wrong.

Neither study proves AI feels anything. Anthropic says directly that “none of this tells us whether language models actually feel anything.” And neither study tested what happens when you, the user, are rude. ChatGPT read traumatic stories. Claude faced repeated failure.

There’s even a small Penn State preprint that found very rude prompts scored slightly higher on ChatGPT-4o: 84.8% vs. 80.8% for very polite ones. It tested one model and hasn’t been peer-reviewed, but it’s worth knowing.

So the real lesson isn’t “say please.” It’s this: stressful inputs and repeated failure change how AI behaves, and not for the better.

Is Anthropic banning people from cursing at Claude?

No, not for normal frustration. On October 8, Anthropic updated its usage policy to prohibit “sustained and needless abusive or cruel behavior toward our models.” It takes effect November 12.

Anthropic says it applies “only in extreme cases” and “does not apply to common versions of user frustration.” If someone keeps at it, Claude can end the chat, and you start a new one.

What’s changing: how you talk to AI is now a work skill

AI is doing real work now: writing code, running tasks, handling clients. When it’s working for you, how you set it up shapes what you get back.

The people who get the best results aren’t the ones who yell louder or just retry. They’re the ones who give clear instructions, reset when things go sideways and check the work.

Your move: five ways to get better results from AI

1. Be clear, not harsh. Say exactly what you want, what “done” looks like and give it context. Clear instructions beat yelling every time.
2. When it keeps failing, reset. Repeated failure is what triggered desperation in Anthropic’s study. Start a fresh chat with a short summary, or break the task into smaller steps.
3. Let it say “I don’t know.” Anthropic’s own guide to reducing made-up answers recommends it. It makes AI less likely to guess or fake it.
4. Make it prove its work. Ask it to show what it changed and how it checked, so “all tests passed” actually means they passed.
5. Keep your tone calm. Frustration is normal. Cruelty doesn’t get you better results.

Paste this at the start of any chat or project:

Take your time and work step by step. If something isn't working or you're not sure, tell me instead of guessing or working around it. Don't say anything is done or passing unless you've actually checked it. When you finish, show me what you changed and how you verified it.

This gives you leverage: fewer fake “done”s, fewer redo’s and more work you can actually trust.

So, are you nice to your AI? Have you noticed a difference? I want to know.

AI is becoming part of how we all work. The people who learn to work with it, not against it, are the ones who get ahead.