Being Polite to AI: Does It Get Better Answers?

Being polite to AI barely moves the quality of what it gives back, and acting hostile or shouting at it won't reliably improve anything either. What moves the result is how plainly you spell out the task.

People get fed up with chatbots, switch to caps lock and start screaming at the screen. We run prompts through over 100 models at ChatFuse, across Google, Anthropic, OpenAI and the rest, so whatever we tell you about speaking to a model needs to work everywhere. The published papers and the model makers' own guides line up more than you might think.

Does being polite to AI get better answers?

Not in any way you'd feel. Penn State ran the cleanest recent test: 2 researchers there wrote 50 multiple choice items on math, science, and history, put each one into 5 different tones from extra polite to outright rude, and tested each batch 10 times using GPT-4o. Very polite prompts hit 80.8%. Very rude ones hit 84.8%.

So being rude won by 4 points. Hold off before you start trash talking your bot and check what the test covered. Only one model. Questions with a right and a wrong answer, period. It says nothing about how well something was written or whether the reasoning helped. The authors wrote they "do not advocate for the deployment of hostile or toxic interfaces," and saw the gap as evidence that models "remain sensitive to superficial prompt cues."

A study before that cut in the opposite direction. In 2024, Ziqi Yin's group used English, Chinese, and Japanese prompts on older models, GPT-3.5 and Llama 2 70B among them. Their finding: "impolite prompts often result in poor performance, but overly polite language does not guarantee better outcomes." How polite you should be moved around depending on the language.

Why did rude prompts score higher in the Penn State study?

Most likely because the rude prompts were tighter and more straight up, though the study never checked that. Read the actual sentences. The very polite one started with "Can you kindly consider the following problem and provide your answer." The very rude set had "Hey gofer, figure this out." One is a request dressed up in nice words. The other is an order.

That's how we read it, and the paper doesn't say so. It lines up with what the AI companies put in their own guides: tell it straight out what you need.

Does yelling at ChatGPT or threatening it work?

No. The clearest test came out of Wharton's Generative AI Labs in August 2025. Google's founder Sergey Brin had said models "tend to do better if you threaten them," and a tip for a chatbot was already a popular trick, so the research team tried both.

They ran 25 tries per question for each prompt, using GPT-4o, GPT-4o mini, o4-mini, and 2 Gemini Flash models on GPQA and MMLU-Pro, 2 hard benchmarks. Their result: threatening or tipping "generally has no significant effect" on benchmark performance. Individual questions swung hard, up 36 points on one side and down 35 on the other, but there was no way to know beforehand which way any one question would tip.

One odd result came out of a prompt where a mom had cancer. Gemini 2.0 Flash got almost 10 extra points on a single benchmark from that one. The people who wrote the paper put it down to a quirk specific to that model.

Shouting at newer models can make things worse. The prompting guide from Anthropic says Claude Opus 4.5 and 4.6 are "more responsive to the system prompt than previous models," so the kind of prompt used to push older models now sends these ones too far. The fix they list is to "soften any aggressive language." What you used to type as "CRITICAL: You MUST use this tool when..." can now just be "Use this tool when..."

OpenAI says much the same thing. The GPT-5 prompting guide from OpenAI says the model "follows prompt instructions with surgical precision." The same guide also brings up Cursor, the coding tool, where the instruction "Be THOROUGH when gathering information" wound up hurting it, since GPT-5 pulls in context on its own already. Cursor fixed that with softer wording. Older versions needed capitals and MUST as a workaround because they skipped instructions. Today's models follow instructions closely, so that old workaround shoves them further than you want.

Should you talk to AI like a human?

Yes. Treat it like a sharp new hire, someone smart but missing your daily context. Anthropic's guide puts it as follows: "Think of Claude as a brilliant but new employee who lacks context on your norms and workflows. The more precisely you explain what you want, the better the result."

That check is a good one. Hand your prompt to someone who doesn't know the task and see if they can carry it out. "If they'd be confused, Claude will be too." This test has nothing to do with Claude in particular. If a new person would be lost reading it, something's missing, and a model can't invent what you didn't tell it.

So give it full sentences and a reason. Talk to it the way you'd talk to a person. Skip the habit of snapping at it, and stop treating the prompt box like a search field. "Marketing email product launch" is what you'd type into search. Give it 3 sentences. Tell the model who you're emailing and what action you want from them, and you'll get a lot more out of it.

What do OpenAI, Anthropic and Google say makes a good prompt?

All 3 guides lean on the same basic rule: be direct about what you're after, and include the background it can't work out alone. Each of them pushes a different habit after that.

Anthropic's trick about giving the reason is the one worth borrowing. "NEVER use ellipses" just tells Claude what not to do. Explain that the answer gets read aloud by a text to speech program that can't say ellipses, and now the rule comes with a why. Anthropic puts it this way: "Claude can generalize from the explanation." Claude can take that why and carry it over to cases you never spelled out.

Google talks the most plainly about examples. Google's guide tells you to put a handful of examples in every prompt, and adds that skipping them tends to hurt performance. Its Gemini 3 advice adds: "State your goal clearly and concisely. Avoid unnecessary or overly persuasive language." A threat or a bribe is persuasive language.

How do you write a prompt that gets the best answer?

Begin by naming the task and the person you're doing it for. Then say what a strong answer would look like, include an example when you have one, and name any limits like how long it should be or how to sound.

On top of that, some habits help. Give the model stuff it can't know on its own: who's going to read it, what happened earlier, what you've already attempted. If a rule is important, tell it why the rule exists. Read once through your instructions and look for contradictions, because OpenAI says "contradictory or vague instructions can be more damaging to GPT-5 than to other models." When you aren't sure what you want, have the model ask you questions before it starts.

Why does the AI keep getting it wrong when I repeat myself?

Because saying the same thing again, only louder, doesn't hand the model anything fresh. If the first try missed, the prompt was short on something. Typing it in all caps doesn't supply the missing piece.

There's another issue too. Most chat apps send the whole history back on every new message, so all the bad attempts stay in front of the model every single turn. Two things tend to help more. Call out the exact miss, like "too formal, and you dropped the refund amount." Or wipe the chat and start over with a first prompt that bakes in what you just learned.

Does the same prompt work on every model?

For the most part, yes. The simple stuff crosses over from one model to the next. Small details won't always survive the trip. In the study, Yin found the proper level of politeness changed with the language. Anthropic's caution on harsh wording is aimed at particular Claude models, and the OpenAI note is aimed at GPT-5.

That's the problem we sit with all the time at ChatFuse. Our Orchestrator needs less than 150 milliseconds to read a prompt and pass it to whichever model fits it best. It picks from over 100 choices made by OpenAI, Anthropic, Google, Meta, and others. We spell out how that sorting happens in AI prompt classification. One straightforward prompt is enough, and you won't need to memorize the little differences between labs. If you'd rather pick the model yourself, you can switch AI models mid conversation without breaking the thread.

The stuff you'd normally retype can live in one spot. ChatFuse Memory hangs onto the useful details and preferences you'll want in later chats, and we cover that in persistent AI memory. A ChatFuse persona keeps standing rules for how it sounds, what to pay attention to, and how to behave. You won't get the same output from every model, but you won't start each chat from blank either. ChatFuse Agents are on the way and work off that same principle: say the job in ordinary words, then check the steps once before it can run.

Is it OK to say please and thank you to AI?

You can. It doesn't push the quality down enough to notice. In the Penn State test, the very polite and neutral prompts ended up 1.4 points apart. If your writing already leans polite, keep that.

Does writing in all caps make AI pay more attention?

With today's models, all caps can make them focus too hard. Anthropic's guidance to builders is to dial back strong words such as "CRITICAL" and "MUST," since newer Claude models overreact to them now. Hold the loudest words for a single rule that cannot move.

Do longer prompts get better answers?

Only if those extra words add information. A paragraph that covers the audience, the aim, and an example does help. A paragraph that pleads does not. Wharton's team told people to use plain, clear directions so the model doesn't get mixed up.

If you'd like to try this with Claude, Gemini, and GPT without paying 3 labs separately, ChatFuse packs more than 100 different models into one $20 a month plan, with a free trial for 7 days. Start your free trial or see ChatFuse pricing.