A new study from Pennsylvania University has revealed that the tone in which users interact with large language models (LLMs) significantly impacts the models’ response quality and efficiency. Researchers examined how responses varied across seven different tones, finding that rude prompts resulted in shorter and more accurate replies, particularly for ChatGPT-4o, which registered 89.04% accuracy with an average of 223 tokens. In contrast, the neutral tone for the Gemini 2.5 Flash Lite model provided 88.25% accuracy but required 1,222 tokens. The study underscores the importance of tone choice, suggesting it is not merely a user experience consideration but also a critical factor for the reliability and cost-effectiveness of LLM outputs, prompting developers to standardize and benchmark prompt styles across models.

arxiv: arXiv is the primary open-access platform for sharing preprints in computer science, physics, and related disciplines. The tone-dependent inference cost paper was uploaded there with identifier 2607.23915. It enables rapid dissemination of AI research findings to the broader technical community.
ChatGPT-4o: ChatGPT-4o is OpenAI’s multimodal language model optimized for conversational interactions and complex reasoning. The study found it delivered peak accuracy and minimal token output under rude prompting compared with other tones tested. This positions tone selection as a relevant factor for applications built around the model.
Gemini 2.5 Flash Lite: Gemini 2.5 Flash Lite is Google’s efficient language model variant built for speed and lower resource demands. Research showed neutral prompting produced the strongest accuracy and shortest outputs for this model, unlike patterns observed elsewhere. The results underscore how model architecture shapes sensitivity to input phrasing.
Pennsylvania University: University of Pennsylvania is an Ivy League research institution with strong programs in computer science and artificial intelligence. Its researchers authored the paper investigating how prompt tone influences large language model behavior. The work was posted on arXiv and focuses on practical implications for AI system design.

Prompt Design: Tone selection in user inputs to large language models can affect both output reliability and computational demands in production settings.
System Benchmarking: Developers are advised to test and standardize prompt styles across different models rather than treating them as simple user interface preferences.