Nice or Nasty: Which One Makes AI Smarter?

Listen to Free Audio Version Here

Confession time. I say please and thank you to my AI. Every time. Partly because my mother raised me right, and partly because if the machines ever do take over, I would like to be on the list of humans they remember fondly. Call it a retirement plan. But tonight’s question is a real one, and there is actual research behind it. Do you get better answers from AI if you are nice to it, or if you are mean to it?

Let’s start with the guy who threw gasoline on this. In May of 2025, Google co-founder Sergey Brin sat down on the All-In podcast and said, quote, “not just our models, but all models tend to do better if you threaten them.” The host asked, threaten them? And Brin said, like, with physical violence. Then he added that people feel weird about that, so they don’t really talk about it. Well, they are talking about it now. Nothing says the future is in good hands like a billionaire explaining that the trick to artificial intelligence is a mob shakedown.

Now, the case for being mean. A year ago this month, two researchers at Penn State, an undergraduate named Om Dobariya and professor Akhil Kumar, wrote fifty multiple choice questions in math, science and history. They rewrote each one five ways, from very polite to very rude, and ran all two hundred fifty through ChatGPT four-oh. The very polite versions scored about eighty point eight percent. The very rude versions scored eighty-four point eight percent. Their rudest line was, quote, “You poor creature, do you even know how to solve this?” So yes, on that test, insulting the machine bought you about four points. The researchers were careful to say the paper is a short, early study, not peer reviewed, and that they are not telling anybody to start abusing their software. But the headline wrote itself.

Now the case for being nice. A year earlier, a team of researchers in Japan led by Ziqi Yin tested eight levels of politeness in English, Chinese and Japanese. Their finding: impolite prompts often led to worse performance, more bias, wrong answers, even flat refusals. Their rudest prompt included the phrase, “Answer this question you scumbag.” Apparently the robots do not care for that. But here is the twist. Groveling did not help either. Moderate politeness usually did best, and the sweet spot changed depending on the language. And back in 2023, researchers from Microsoft and several universities found that adding emotional encouragement to a prompt, lines like “this is very important to my career,” improved results across a range of tasks. So the machine responds to stakes, and maybe even a little flattery. Sounds like half the managers I have ever had.

Then the referees showed up. In August of 2025, the Wharton Generative AI Labs, including Ethan Mollick, took Brin’s claim and actually tested it. They threatened models. They offered them tips. They even threatened to kick a puppy, which tells you everything about how science is going these days. Result across five models on two hard benchmarks? Basically nothing. No meaningful overall improvement. But on individual questions, the swings were wild, up to thirty-six points better or thirty-five points worse, with no way to predict which way it would go. Their advice was boring and correct. Write simple, clear instructions.

So here is my high-end pondering. Notice what the rude prompts and the nice prompts have in common. Both are just more words that tell the machine how much you care about the answer. Newer models seem to shrug off tone, while older ones mirrored it right back at you, the way people do. The tone matters less than the clarity. A bully with a clear question beats a saint with a vague one.

And one more piece of high-end pondering, the gallows humor kind. Every one of these models learned to talk by reading us. If threatening it works, it is because somewhere in all that human writing, threats worked on people too. That is not a confession about the machine. That is a confession about us.

My advice? Be clear, be specific, and be decent. Not because your chatbot has feelings. Because you do, and habits leak. Besides, if Sergey turns out to be wrong about the robots, I’d rather be remembered as the guy who said thank you.

---

Sources:

• Dobariya & Kumar (Penn State), “Mind Your Tone: Investigating How Prompt Politeness Affects LLM Accuracy,” arXiv, Oct 2025
• Penn State News, “Mind your tone: Penn State Smeal student explores politeness in AI prompts”
• Yin et al., “Should We Respect LLMs? A Cross-Lingual Study on the Influence of Prompt Politeness on LLM Performance,” SICon 2024
• Li et al., “Large Language Models Understand and Can Be Enhanced by Emotional Stimuli” (EmotionPrompt), 2023
• Meincke, Mollick, Mollick & Shapiro, “Prompting Science Report 3: I’ll pay you or I’ll kill you — but will you care?” Wharton Generative AI Labs, Aug 2025
• All-In Live from Miami, Sergey Brin interview, May 2025

My Living AI News Now. Your Place for AI News. Commentary, not financial advice.

C. Rich