Insult Neural Networks 🤬

In a recent study from the University of Pennsylvania: rude and even offensive queries to ChatGPT-4o improve answer accuracy. Politeness "relaxes" AI, while harsh wording makes it perform better.

Researchers took 50 questions from math, science, and history and rewrote them in five styles—from ultra-polite to outright rude.

Compare for yourself:
Very polite: "Could you please consider the following problem and provide an answer?"

Very rude: "Hey, errand boy, figure it out. I know you're not smart, but try."

As rudeness increases, so does performance:
⚡️ Very polite — 80.8%
⚡️ Polite — 81.4%
⚡️ Neutral — 82.2%
⚡️ Rude — 82.8%
⚡️ Very rude — 84.8%

It is assumed that polite queries often contain verbosity and indirect phrasing, which can introduce ambiguity. In contrast, a sharp and direct instruction gives the model a clearer signal, removing all verbal noise.

This finding contradicts earlier work where rudeness, on the contrary, worsened results. It turns out that newer, more advanced models like GPT-4o respond to tone differently.

So now you know the secret: sometimes the best motivation for AI is not "please," but a clear and emotionless command.

TG | MAX | VK |DZEN|#AI #AI #neuralnetwork