Research examines impact of prompt tone on AI performance

Recent studies are investigating whether the tone of user prompts—ranging from polite to rude—affects the quality of responses from generative artificial intelligence models like ChatGPT.

Findings across different research institutions remain inconclusive and vary by model, language, and task. A study from Pennsylvania State University involving GPT-4o found that very rude prompts achieved an 84.8% accuracy rate, compared to 80.8% for very polite prompts, though researchers noted the study was limited to a single model.

In contrast, researchers from Waseda University and the RIKEN institute in Japan found that impolite commands often worsened performance, while excessive politeness did not guarantee better results. Their findings suggested that the ideal level of courtesy varies depending on the language used. Additionally, research from the Wharton School involving GPT-4o and GPT-4o mini found no significant evidence that specific prefixes, such as “please respond” or “I order you to respond,” altered performance outcomes.