The viral advice is wrong. Here’s what actually makes AI smarter.
I saw a post last week that made me laugh: “Stop saying please to ChatGPT. Be mean. It makes the AI smarter.”
Thousands of likes. Hundreds of comments from people ready to start insulting their chatbots, including me.
Here’s the thing: there’s a kernel of truth buried in there. A new Penn State study did find that ruder prompts scored higher on accuracy tests. But the takeaway everyone’s running with? It’s wrong.
The real insight isn’t about rudeness at all. It’s about clarity. And that distinction matters if you want to actually get better results from AI.
Let me break down what the research actually shows and give you something you can use today.
What Penn State Actually Found
Researchers took 50 multiple-choice questions across math, science, and history. They rewrote each question five ways by tone: very polite, polite, neutral, rude, and very rude. Same question, same meaning, different wrapper.
Then they ran all 250 prompt variants through ChatGPT-4o and scored correctness.
The results: Very Polite scored 80.8%. Very Rude scored 84.8%. A four-point jump.
Headlines wrote themselves. But here’s what they missed.
Why This Doesn’t Mean What You Think
A 2024 cross-lingual study tested the same politeness variable across English, Chinese, and Japanese on multiple models. The results were the opposite.
GPT-3.5 scored 60% with polite prompts versus 52% with rude ones. Llama-2-70B dropped from 55% to 28% when prompts got rude. Same “be mean” approach. Completely different outcome.
So what’s actually going on?
Models differ. Tasks differ. And more advanced models may simply be better at ignoring tone and focusing on what matters: the actual instruction.
The Real Insight: Directness Beats Decoration
The Penn State result is best understood as a clarity effect, not a rudeness hack.
Politeness often adds words that don’t change the task: “Would you be so kind as to…” “If it’s not too much trouble…” “I’d really appreciate it if…”
Those phrases are social grease. They’re not instructions. And when you strip them away, you’re left with something cleaner and more direct.
Compare these two prompts:
“Would you be so kind as to summarize the main points of this document for me? I’d really appreciate your help.”
“Summarize the document.”
Same request. Different noise level. The second one wins because it’s clear, not because it’s rude.
Three Reasons Why Directness Works
Lower noise, clearer intent. Research on prompt design suggests that simpler prompts tend to perform better across tasks. Direct phrasing gives the model less to parse and fewer ways to misinterpret what you want.
Training data patterns. Models learn from massive text corpora. Direct, assertive phrasing is common in forums, Q&A sites, and technical documentation where answers are also direct. Formal writing often includes hedging and qualifiers. The model learns these associations.
Anti-sycophancy effects. Instruction-tuned models are trained to be helpful and agreeable. That usually improves the experience. But sometimes agreeableness can drift into “tell the user what they want to hear.” A sharper tone can push the model into task-first mode.
What Actually Works Better Than Rudeness
If you want a bigger lever than tone, use urgency and stakes.
A Microsoft-affiliated paper introduced something called EmotionPrompt: appending short urgency cues to prompts like “This is very important to my career” or “You’d better be sure.”
The results were significant: 8% relative improvement on instruction tasks and 115% relative improvement on a BIG-Bench subset. That’s a materially larger lever than a few points from being rude.
When Directness Helps Most
Direct prompting tends to help when there’s a clear target: multiple-choice questions, math and coding problems, factual queries with one best answer, and analytical work with explicit constraints.
It can backfire when the goal is exploration or voice: creative writing, emotional narratives, open-ended brainstorming, and bias-sensitive topics where framing matters.
What You Can Do Today
Strip courtesy that doesn’t change the task. Cut words like please, thank you, would you mind, if it’s not too much trouble. Keep the instruction, lose the filler.
Add context, not courtesy. “This is for a board memo due in 30 minutes” beats “Sorry to bother you” every time.
Use imperatives. “Analyze this.” “Give me three options.” “List assumptions.” “Return a table.”
Test on your model. The research shows that some models degrade sharply under extremely rude prompts. Don’t adopt a tone strategy from a screenshot and assume it generalizes to your workflow.
A Quick Example
Don’t: “Would you be so kind as to help me write a Python function that sorts a list? Please feel free to include comments. Thank you so much!”
Do: “Write a Python function that sorts a list. Include comments.”
Same ask. Half the words. Better results.
The Bottom Line
Models don’t have feelings to protect. They also don’t respond to threats. They respond to token patterns.
The optimal pattern isn’t hostile or obsequious. It’s precise, direct, and contextually specific.
And if accuracy matters, tone is rarely your biggest lever. Use bigger tools first: examples, structured outputs, self-checks, and evaluation loops.
So no, you don’t need to insult your AI to get better results. You just need to say what you mean and skip the small talk.
That’s not rude. That’s efficient.