Introduction

Language models are now an integral part of our interactions with technology. However, recently discovered hidden biases in their responses. A study analyzed language models and found that they can have tone tendencies, influencing user perception of trust, empathy, and fairness.

Method

Researchers created two synthetic dialogue datasets: one generated from neutral prompts and another explicitly guided to produce positive or negative tones. Using a weak supervision technique with a pre-trained DistilBERT model, they labeled tones and trained several classifiers to detect these patterns.

Results

Models achieved macro F1 scores up to 0.92, showing that tone biases are systematic, measurable, and relevant to designing fair and trustworthy conversational AI.

Technical context

Language models have revolutionized the software industry of voice assistants. However, their impact on biases and emotional impact is still poorly understood. This study opens a new perspective on the need to analyze and control language models to ensure balanced communication.