Bias Beneath the Tone: Empirical Characterisation of Tone Bias in LLM-Driven UX Systems

This research paper investigates the subtle yet significant issue of tone bias in Large Language Models (LLMs) used in conversational UX systems. The study highlights that even when prompted for neutral responses, LLMs can exhibit consistent tonal skews, potentially impacting user perception of trust and fairness. The methodology involves creating synthetic dialogue datasets and employing tone classification models to detect these biases. The high F1 scores achieved by ensemble models demonstrate the systematic and measurable nature of tone bias. This research is crucial for designing more ethical and trustworthy conversational AI systems, emphasizing the need for careful consideration of tonal nuances in LLM outputs.

• •LLMs exhibit tone bias even when prompted for neutral responses.
• •Tone bias can impact user perception of trust and fairness.
• •Tone bias is systematic and measurable using tone classification models.