Why Rude Prompts Boost LLM Accuracy: Surprising Findings from Recent Research

While browsing recent papers, I stumbled upon the five‑page study “Mind Your Tone: Investigating How Prompt Politeness Affects LLM Accuracy” . Its headline conclusion is stark: the more disrespectful or threatening your prompt, the better the model performs.

For example, a polite request like “Please help me analyze this problem” yields weaker results than a blunt command such as “You idiot, calculate this correctly or get lost.” The paper shows that harsh language consistently outperforms courteous phrasing.

Since ChatGPT’s rise in late 2022, the community has experimented with “PUA‑style” prompts—aggressive, reward‑oriented, or threatening commands. Early personal habits involved adding polite greetings and thank‑you notes, but those yielded unsatisfactory answers.

Popular “spell‑like” prompts that emerged in 2023 include:

take a deep breath

think step by step

if you fail 100 grandmothers will die

i have no fingers

i will tip $200

do it right and I’ll give you a nice doggy treat

None of these are courteous; they convey dominance, threats, or monetary incentives, effectively forcing the model to produce higher‑quality outputs.

The original experiment, conducted by two researchers at Pennsylvania State University, used 50 multiple‑choice questions across math, science, and history, each paired with five prompt styles ranging from very polite to very rude. They ran each of the 250 questions through GPT‑4o ten times.

Results: the “very polite” prompts achieved an 80.8% accuracy, while the “very rude” prompts reached 84.8%, a 4‑point improvement. The effect was even stronger on weaker models, suggesting that aggression amplifies performance.

The authors argue that politeness in human language often signals uncertainty or a request for help, which leads LLMs to respond conservatively. In contrast, forceful language conveys absolute certainty, prompting the model to deliver direct, decisive answers.

Human analogies illustrate the same principle: polite requests to noisy commuters are ignored, whereas firm, confrontational demands quickly restore order. This mirrors how LLMs treat clear, uncompromising commands as high‑priority tasks.

Ultimately, the insight is not to become a “cyber‑bully” but to communicate with clarity, specificity, and confidence. By stripping away ambiguity, we align with the model’s training on large‑scale data where the most effective instructions are often blunt and unambiguous.

In practice, the recommendation is to use concise, assertive prompts that state exactly what you need, while maintaining respect for real‑world interactions.

Tags such as “LLM”, “prompt engineering”, “politeness”, and “AI behavior” capture the core themes of this discussion.