The paper „Reading Between the Laughs: A Human-Referenced Audio Evaluation of MLLMs for Social Robotics“ by Sahan Hatemo, Katharina Kühne, and Oliver Bendel has been accepted at ICSR + Art 2026. In this work, the researchers investigated whether today’s leading AI models can distinguish genuine from non-genuine laughter based solely on audio signals. The results revealed striking differences in model behavior: OpenAI systems showed a strong tendency to interpret most laughter as genuine, while Gemini models were generally more skeptical. Despite these contrasting biases, several models performed significantly better than chance, with Gemini 2.5 Pro achieving the strongest overall results. Their analysis also demonstrated that less capable models often relied on superficial cues such as pitch, disproportionately labeling higher-pitched laughter as less authentic, whereas the top-performing model appeared to focus on more sophisticated voice quality features, suggesting a deeper understanding of laughter authenticity. These findings highlight the growing potential of multimodal large language models in social robotics, where accurately interpreting subtle social signals like laughter could play an important role in trust, communication, and relationship building between humans and robots. The 18th International Conference on Social Robotics will take place in London, UK, from 1-4 July 2026. ICSR is the leading international forum that brings together researchers, academics, and industry professionals from across disciplines to advance the field of social robotics.

Abb.: Reading between the laughs