AI mimics human emotions in groundbreaking therapy research study

A new study in The Lancet Digital Health suggests large language models can mimic human-like emotional responses. Researchers tested six advanced AI systems with scenarios designed to trigger emotions such as fear, sadness, and anger. The findings open possibilities for early-stage therapy research but come with clear limitations. The study exposed AI models to emotionally charged situations, including stress and disgust. When faced with sadness-inducing scenarios, the models displayed a negativity bias, mirroring patterns observed in humans with low mood. Their 'emotional scores' rose in response, showing an ability to adapt outputs to context.

A mindfulness-style breathing exercise was then introduced. The models’ emotional scores dropped, indicating they could simulate a calming effect. Researchers describe the work as a proof-of-concept, demonstrating that AI can reproduce certain human-like response patterns under controlled conditions.

The team argues these models could serve as a flexible tool for testing talk-based therapy ideas early on. They stress, however, that AI does not experience genuine emotions. Small changes in wording or context can also lead to inconsistent or fabricated responses. The study highlights potential for AI in supporting early therapy research. It does not suggest AI can replace human therapists or feel emotions. Experts warn against overinterpreting the results, as the models only simulate responses rather than truly experiencing them.