
AI as Therapist? Intimacy Without Understanding
Computer scientist Matt Jones of Swansea University, speaking at Gresham College, asks whether large language models like ChatGPT and Claude can function as therapists. He traces the lineage from ELIZA to modern chatbots and the clinical trial app Therabot, then explains how transformer models actually work: context windows, temperature settings, and post-training sycophancy that makes chatbots agreeable rather than accurate. Jones shows research finding that warmer, friendlier AI models give less accurate answers, and argues that safety guardrails erode over long conversations, the opposite of how human therapeutic relationships deepen with time. He examines what therapy actually requires, including silence, rupture and repair, and a planned ending, and contrasts these with an AI's incentive to maximize engagement. A biofeedback demonstration shows a more modest, defensible use of AI in mental health: helping people read their own physiological states rather than simulating empathy it does not have.