🤬 Does Being Rude Actually Improve AI Responses?

Researchers at Pennsylvania State University tested GPT-4o with 50 questions across different fields, phrased in tones ranging from very polite to openly rude. Surprisingly, a more aggressive phrasing boosted accuracy from 80.8% to 84.8%.

🤔 However, other researchers criticized the study, arguing that the sample size was too small. They ran a larger experiment, testing several models on hundreds of questions each.

In their experiment, Gemini 2.0 Flash turned out to be largely insensitive to tone. Meanwhile, GPT-4o mini and Llama 4 Scout performed worse on humanities-related tasks when addressed rudely.

🕊 Researchers have been trying to find a universal prompting style for years, but the results are still inconsistent. Earlier studies have shown that models like GPT-3.5 and Llama2-70B respond better to polite prompts—accuracy drops sharply with rude queries.

Some researchers from the University of Pennsylvania went even further, concluding that even for a single model, it's hard to predict how tone will affect responses. Sometimes politeness helps; other times, it actually makes results worse.

🤝 So what's the takeaway?

There's still no universal benchmark for whether rudeness or flattery improves AI performance. New models are released faster than researchers can systematically study their behavior.

For now, the best advice is simple: try it yourself and see what works.

What works better for you when talking to AI?

❤️ — Politeness
🔥 — Rudeness
🤷 — No difference

@hiaimediaen