Proceedings of the Annual Meeting of the Cognitive Science Society

Many people treat large language models (LLMs) like counsellors, seeking advice on social and emotional issues. This could improve quality of life for people who lack access to mental healthcare, but it also comes with serious risks. We recently got the opportunity to analyze written Chinese conversations between hundreds of human–human client–counsellor pairs and thousands of human–AI client–counsellor pairs. By analyzing linguistic and semantic features of those conversations, we provide evidence that human and AI counsellors share many qualities, such as writing predictable messages and expressing positive sentiments, and that human clients align with AI counsellors more than with human counsellors, including when AI behaviour diverges from human counsellor behaviour. This study contributes sorely needed empirical descriptions of real-world counselling sessions with chatbots, which can help researchers at the intersection of AI and psychology understand whether and how LLMs should play a role in mental healthcare.