The tone of a question influences generative AI responses, but the effect remains uncertain

What do studies say about tone and performance?

Penn State: Rude commands yielded better results

A study by Pennsylvania State University researchers Om Dobariya and Akhil Kumar used 50 base questions rewritten in five tones, ranging from very polite to very rude. GPT‑4o responded to 250 variations. A polite request was phrased as, “Can you kindly analyze the following problem?” while the rude version said, “Hey, helper, solve this.” The very rude tone achieved 84.8% accuracy, compared to 80.8% for the very polite tone. The article has not yet been peer-reviewed and has been criticized for using a single model and a limited number of trials.

Waseda and RIKEN: Rudeness tends to worsen outcomes

In 2024, researchers from Waseda University and Japan’s RIKEN institute evaluated GPT‑3.5 and GPT‑4 in English, Chinese, and Japanese. Rude requests often reduced performance, but excessive politeness did not guarantee better results. The ideal level of courtesy varied by language.

Wharton: The effect appears to depend on context

In 2025, Ethan Mollick’s team at Wharton compared prefixes such as “please answer” and “I command you to answer” in GPT‑4o and GPT‑4o mini. Each condition was tested 100 times on 198 advanced questions. The authors found no significant difference between conditions, except for one exception in the smaller model. They also noted that small changes in text can lead to large variations in the same model.

Other models show distinct behaviors

A test involving GPT, Gemini, and Llama revealed that sensitivity to tone varies by model and by the area being evaluated. Gemini proved to be comparatively insensitive to changes in tone.

How to apply this in the workplace

The Penn State and Wharton experiments used multiple-choice and academic questions, not everyday tasks like writing emails or summarizing contracts. Therefore, the results cannot be automatically applied to daily professional life. One suggested practice is to ask an analyst to formulate the same request in two ways—one polite and one direct—and compare the responses before setting a standard for the team. Repeating the test is useful, as even an identical question can yield different results on different occasions.

What remains uncertain

Research is ongoing, and findings may change with new generations of models. So far, evidence suggests that tone can affect performance in some cases, but the effect is inconsistent and depends on the model, language, and specific task.