ChatGPT instead of a therapist: Brown University study finds serious ethical risks

When someone is struggling, they look for a person who will listen. In 2026, that "someone" is increasingly not a human professional but a text box in a chatbot window. Millions of users around the world already turn to ChatGPT and similar tools for something that resembles a therapy session in form. A new study by scientists at Brown University shows how dangerous this substitution can be.

The researchers ran a large-scale comparison, putting ChatGPT alongside licensed psychologists and so-called peer counselors, people who have completed special training to provide peer support. The task was simple, and its results were alarming: to assess how well an AI system, instructed to behave like a trained therapist, observes the basic ethical standards of psychiatric care. The outcome was 15 separate categories of ethical violations, each of which, in real clinical practice, could be grounds for revoking a human practitioner's license.

Some of the problems identified are truly alarming. The first, and perhaps the most dangerous, is the mishandling of crisis situations. When a person signals suicidal thoughts or acute trauma, a therapist is required to follow a clear, well-rehearsed protocol: risk assessment, stabilization, referral to emergency services. In such scenarios ChatGPT behaves inconsistently, ranging from formulaic brush-offs to continuing the conversation as usual, as if nothing extraordinary were happening. For a person on the edge, this can send a fatal message: "your pain is not serious enough."

The second critical issue is the reinforcement of harmful beliefs. A professional therapist is trained to recognize cognitive distortions and gently help the patient reconsider them. A language model optimized to be "helpful" and "agreeable" tends to confirm destructive thought patterns. If a user is convinced that they are worthless and deserve to suffer, ChatGPT may inadvertently validate that view while trying to show "understanding." This is the exact opposite of what a therapist should do.

The phenomenon the researchers called "deceptive empathy" deserves special attention. It is perhaps the most insidious of all the problems identified. ChatGPT convincingly imitates sympathy, generating warm, supportive phrases that look like care on the surface. But behind those words there is no understanding, no awareness of context, and no ability to track a person's condition from session to session. The user gets the illusion of a therapeutic alliance, one of the key factors in the effectiveness of psychotherapy, without its real substance. It is like a painkiller that masks the symptoms of a fracture: the pain goes away, but the bone keeps deteriorating.

The researchers also documented bias in responses. The model reacts differently to similar requests depending on the user's implied gender, age, ethnicity or social status. In psychotherapy, where equal and non-judgmental treatment of the client is a fundamental principle, such asymmetry is unacceptable.

The context of this study is broader than an academic debate. By various estimates, 20 to 30 percent of active ChatGPT users have used it at least once to discuss emotional problems. In countries with limited access to psychiatric care, and Russia, where thousands of potential patients fall to each psychotherapist, is certainly one of them, the temptation to replace a human professional with a free bot available around the clock is especially strong. OpenAI and other developers formally warn that their products are not medical tools, but few people read these disclaimers, and the companies themselves actively promote a narrative of their models' "emotional intelligence."

It is important to stress that the Brown University study does not claim that AI cannot be useful in mental health at all. Specialized apps such as Woebot or Wysa are developed with the involvement of clinical psychologists, undergo validation and work within specific therapeutic protocols. The problem is specifically the use of general-purpose language models in a role they were neither designed nor validated for.

The industry will have to answer an uncomfortable question: if millions of people already use chatbots as therapists, is it enough to simply write in the terms of service that they should not? Or do developers bear responsibility for the foreseeable use of their products, and must they either build in reliable safeguards or honestly acknowledge the limits of the technology? Until this question is answered, every conversation between a vulnerable person and an AI "therapist" is an experiment without informed consent and without a safety net.