This is not a provocation or a bar joke: responding badly to the artificial intelligence could make it… more precise. This is supported by a study by the University of Pennsylvania, published on October 25 on arXivwhich put the prompt language to the test to understand how much the tone may influence the effectiveness of ChatGPT responses.
Yes, you understood correctly: the ruder you are, the better the AI seems to work. But all that glitters is not gold.
Test on 250 prompts
The research team took 50 multiple choice questionsspanning mathematics, science and history. Each question was reformulated in five different tones: very polite, kind, neutral, rude and very rude.
The result? Prompts with a “very rude” tone have reached a84.8% accuracyagainst the80.8% “very kind” prompts. A gap which, for a model like ChatGPT-4o, is anything but negligible.
Here is a concrete example:
The disparaging language, however objectionable, led to responses more correct and faster. Researchers talk about a “surprising correlation” between rudeness and accuracy. But be careful: the explanation is not at all simple.
Behind the most accurate answers there are also risks
The authors of the study are keen not to send the wrong messages: . The study is limited: it was tested only on ChatGPT-4o, with a small sample, and in English. So, .
Furthermore, there is an important precedent: a 2024 study (Yin et al.) demonstrated the exact opposite. Older models, such as ChatGPT-3.5 and LLaMA2, they go into crisis if you treat them badly: they give worse, more confusing answers, or even refuse to answer.
And then there is an ethical problem: according to another research (Naderi et al., 2025), in the medical field, the use of emotionally manipulative prompts (like “the patient is dying, help!”) pushes the AI to respond with too much securityeven when it does not have sufficient basis. Dangerous behavior, especially in clinical or legal settings.
In short, rudeness can work, but . Not only because the effect is not guaranteed, but also because communicating with empathy and respect is a value that must be cultivated, regardless of the interlocutor.
Treating a linguistic model badly to “make him give his best” is a bit like yelling at a barista for having the best coffee: it may happen once, but in the long run it destroys the relationship, even with technology.
And if it really doesn’t work, remember: behind every AI there is a human team that designed it, and maybe it’s worth approaching with kindness… even when it comes to code.
