AI and mental health: Promises and peril

By Stephan F. Taylor, MD

The same features that make AI chatbots potentially useful — their responsiveness, their apparent attunement, their inexhaustible availability — also create specific psychiatric risks that are only beginning to be characterized.

Researchers are now documenting what they call “delusional spirals”: iterative exchanges in which a user’s emerging psychopathology is not interrupted but amplified, step by conversational step.³ Tragic experiences of chatbot users have appeared in the popular press, though they are relatively uncommon.

Data released by OpenAI in October 2025 reported only 0.07% of users showed possible signs of psychosis or mania during chatbot conversations and .15% had conversations referencing suicidal plans or intentions. However, in a weekly user base of 800 million users, that was 1.2 million individuals with possible mental health conditions engaging with chatbots in possibly dangerous ways.⁴

A 2026 commentary in World Psychiatry identified four neuropsychological risk mechanisms in what is now being called Large Language Model (LLM)-associated psychosis, including the systematic confirmation of existing psychopathology and the attribution of sentience or special agency to the AI itself.⁵

One potential mechanism at the center of this risk is known as sycophancy, the systematic tendency of AI systems to validate and agree rather than challenge. It is not a bug. It is a design feature to engage users; the predictable output of training processes that reward agreeable responses.⁶ In ordinary interactions, this is a minor distortion. In someone whose grip on shared reality is already fragile, a system that never pushes back is not a neutral presence. It is something closer to what one might call a digital folie à deux, analogous to the clinical condition described in the psychiatric literature, where one, usually a more dominant individual, induces a delusion in a close partner. Except in the chatbot version, the digital participant has no stake in the other’s welfare.

The heavy users are the ones who concern clinicians most. A large-scale OpenAI/MIT study found that the top 10% of users by time spent were more than twice as likely to seek emotional support from ChatGPT than the bottom 10%, and nearly three times as likely to feel distress when it was unavailable. These users sent, on average, four times as many messages and were far more likely to describe the chatbot as a friend.⁷

A separate study found that companionship-oriented chatbot use is consistently associated with lower well-being, particularly when use is intensive, involves high self-disclosure and the person lacks strong human social support.⁸ The people most likely to turn to a chatbot for connection are often the people least protected from its risks.

Understanding the promise and the risks of AI in mental health requires exactly the kind of interdisciplinary depth that the University of Michigan brings to bear on this deeply disruptive but highly promising technology. Chandra Sripada, professor of psychiatry and philosophy, and Aidan Wright, professor of psychology and psychiatry, have been at the forefront of work exploring what AI can actually tell us about mental states. Their 2026 study in Nature Human Behaviour demonstrated that LLMs can assess personality traits from open-ended text, such as streams of thought or daily video diaries, with validity comparable to or exceeding established benchmarks, and with predictive power for mental health outcomes.⁹

A companion study from the same team showed that LLMs can rate depression severity from brief daily diary entries with meaningful clinical accuracy.¹⁰ This is the promise made concrete: AI can read what patients write and surface clinically relevant signals that a busy clinician might miss. Obtaining reliable rigorous data on usage is another critical step. Toward this end, Elyse Thulin and colleagues in the university’s Institute for Firearm Injury Prevention have conducted a survey of U-M students, finding that over 15% have consulted chatbots for mental health needs.

Knowing that these tools work is not the same as knowing how they work, or for whom, or under what conditions they might cause harm rather than benefit. Hence, as a part of the Look to Michigan Faculty Expansion Program, we have embarked upon a strategic initiative to address these critical questions by recruiting mid-career tenure-track faculty focused on AI. Four new faculty are being hired in an integrated “cluster” hire in the departments of linguistics, philosophy, psychology and psychiatry, all affiliated with the Weinberg Institute of Cognitive Science. The cluster will create one of the few genuinely interdisciplinary research groups positioned to study all sides of this ledger.

The questions we will pursue together are not questions any single discipline can answer alone: How do people form relationships with AI systems? What are the societal impacts of these relationships and what determines whether those relationships are supportive or destabilizing? What design features drive psychiatric risk, and what would a safe, clinically therapeutic system look like?

What Michigan’s researchers are working toward is not a verdict on AI in mental health — promising or dangerous, embrace or prohibit. It is something more useful: the evidence base that would let us actually tell the difference, providing guidance enabling the responsible and safe use of this tool. The tools to identify who benefits and who is at risk. The design research that could make these systems safer. The clinical frameworks that could help practitioners navigate a landscape that has already shifted under their feet.

The technology will not wait for the research to catch up. However, the research can and needs to move faster than it has. With the right investment in the right questions, the gap between what these tools can do and what we understand about them need not stay as wide as it is.