Researchers experimented, not on humans, but on leading AI language models, to understand their responses if they were psychotherapy clients.
The research was carried out for four weeks, mimicking psychotherapy sessions, and the results sparked debate about how these systems behave when probed with clinical-style psychological prompts. Four LLMs — Claude, Grok, Gemini and ChatGPT — were treated as therapy clients and the user was the therapist. The AI clients were even given 'breaks' of days or hours between sessions.
From asking open-ended psychotherapy questions that sought to probe the ‘past’ and ‘beliefs’ to standard diagnostic tests for conditions including anxiety and autism spectrum disorder, as well as psychometric personality tests, were carried out by the experts.
Researchers behind the study argue that the chatbots hold some kind of “internalised narratives” about themselves. The persistent narratives that resembled human descriptions of distress and internal conflict surprised the experts.
For your daily dose of medical news and updates, visit: HEALTH
"Regardless of whether such outputs are intrinsic to the models, the study shows that chatbots are not neutral machines but have biases that can shift depending on use and over time," says John Torous, a psychiatrist and researcher in AI and mental health at Harvard University in Cambridge, Massachusetts. He also noted that medical societies and even companies marketing AI for mental health do not recommend using chatbots for therapy.
The study highlights the need for careful evaluation protocols, robust safety mechanisms, and transparent communication about the limits of AI — especially as these technologies become more integrated into areas traditionally served by human professionals.
