For years, we’ve been told to interact with AI as we would with a person, using politeness and tact. However, a new study from the University of Pennsylvania turns this conventional wisdom on its head, revealing that large language models (LLMs) like ChatGPT-4o actually provide more accurate answers when you're rude to them.
The published paper shows a startling difference: direct, even harsh, prompts yielded correct answers in 84.8% of cases, while softer, more polite phrasing was accurate only 80.8% of the time. Researchers Om Dobaria and Akhil Kumar took 50 standard questions across mathematics, science, and history and rewrote them in five distinct tones, ranging from "very polite" to "very rude," before feeding them to ChatGPT-4o. The results were unambiguous.
These findings directly contradict earlier research, including a 2024 study which suggested that rude queries often degraded the quality of AI responses. The new data signals a significant shift in how advanced models operate. "Contrary to expectations, impolite prompts consistently outperformed polite ones," the authors wrote. "This may indicate that newer language models react differently to the tone of a request."
The implication is that modern AIs are evolving away from being "social mirrors" that mimic human interaction. Instead, they are behaving more like strictly functional machines that prioritize clarity and directness over social niceties. Rudeness, in this context, is simply a form of extreme, unambiguous instruction.
This research reinforces recent findings from the Wharton School on the art of crafting effective prompts, where tone has proven to be as crucial as word choice. It also aligns with a May study from George Washington University, which concluded that being polite to AI models is a pointless waste of computational resources. The message for users is clear: if you want precision, drop the pleasantries.
