Mental health is the area where AI in healthcare gets the most public attention and the least clinical certainty. Chatbots marketed as therapists sit next to FDA-cleared prescription apps, unregulated companion bots, and peer-reviewed clinical trials, often blurred together in coverage as if they were one category. They are not. This piece separates what has been tested in controlled trials, what regulators have actually cleared, and what has caused documented harm.

Why AI Is Entering Mental Health Care Now: the Access Gap Behind the Hype

The pitch for AI mental health tools rests on a real, well-documented supply problem, not a manufactured one. HRSA’s Bureau of Health Workforce reported that 137 million Americans, roughly 40% of the population, lived in a designated Mental Health Professional Shortage Area as of December 2025, up from about 122 million a year earlier. The same data projects a combined shortfall of roughly 99,840 psychologists, 99,780 mental health counselors, 77,050 addiction counselors, and 36,780 adult psychiatrists by 2038. Stanford researchers framed the problem the same way: nearly 50 percent of people who could benefit from therapy are unable to access it. That gap is the reason AI tools have gained traction, but it doesn’t automatically make those tools safe or effective. For broader context on how AI is being adopted across healthcare generally, see our simple guide to what AI in healthcare actually means.

The AI Mental Health Tools Available Today: Regulated vs. Unregulated

Not every product called an “AI mental health tool” belongs in the same category. The table below separates them by what’s actually been tested and what regulatory status, if any, applies.

Tool Type	Example	Regulatory Status	Evidence Base
Prescription digital therapeutic	Rejoyn	FDA-cleared for adjunctive MDD treatment	386-person RCT (Mirai study)
CBT-based wellness chatbot	Woebot	Not FDA-cleared; positioned as wellness app	RCT showing reduced PHQ-9 depression scores
CBT-based wellness chatbot	Wysa	FDA Breakthrough Device Designation for chronic pain program	Multiple peer-reviewed feasibility and RCT studies
General-purpose LLM used off-label	ChatGPT and similar	Not a medical device; no mental health clearance	Flagged in Stanford’s safety evaluation
Social/companion chatbot	Character.AI, Replika	Not regulated as health product	Assessed as unacceptable risk for minors by Common Sense Media/Stanford Brainstorm Lab

FDA-Cleared Prescription Digital Therapeutics: Rejoyn

Rejoyn, developed by Otsuka and Click Therapeutics, became the first prescription digital therapeutic cleared for adjunctive treatment of major depressive disorder symptoms. Its clearance rested on the Mirai trial, a 13-week, multicenter, double-blinded, randomized controlled trial of 386 participants aged 22 to 64 with MDD who were already on antidepressant medication. It is not a conversational chatbot; it delivers structured cognitive emotional training exercises over a six-week program and requires a prescription. Clinicians reviewing the trial data have noted the results, while positive, leave some questions about real-world effect size, since participants entered the trial with only partial remission on their existing antidepressant and scored an average of 28 on the MADRS depression scale at baseline.

Evidence-Based Wellness Apps: Woebot and Wysa

Woebot has the longest published track record. In the original randomized controlled trial, 70 young adults were split between the Woebot chatbot and an information-only control; the Woebot group showed a significant reduction in depression symptoms on the PHQ-9 (F=6.47, P=.01), while the control group did not. Both groups improved on anxiety (GAD-7), and 83% of participants completed the two-to-three week study. It’s worth noting the study’s second author founded Woebot Labs, a financial conflict of interest disclosed in the paper itself.

Wysa’s evidence base is broader across geographies and conditions. An independent, peer-reviewed trial published in JMIR found Wysa effective for managing chronic pain and associated depression and anxiety, performing better than standard orthopedic care and comparably to in-person counseling, a result that supported its FDA Breakthrough Device Designation. A separate review of that same orthopedic trial found 61 participants with a median age of 55 completed two months of app use, with an 84% retention rate and a clinically meaningful decrease in depression scores among high-frequency users. High engagement, not just having the app installed, appears to matter for outcomes, a pattern echoed in our broader look at what AI patient engagement tools actually do.

Where AI in Mental Health Falls Short: Documented Failures and Open Risks

This is the section clinicians and health IT leaders should read most carefully, because the failures here are not hypothetical.

The Stanford Findings on Stigma and Crisis Response

In 2025, Stanford researchers evaluated five chatbots marketed for accessible therapy against established criteria for what makes a good human therapist. The study found the chatbots could deliver stigmatizing, inappropriate, or unhelpful responses, especially around complex or severe conditions, showing increased stigma toward alcohol dependence and schizophrenia compared to depression. Notably, the paper’s lead author found that bigger, newer models showed no less stigma than older ones. In a second experiment using real therapy transcripts, chatbots sometimes failed to address symptoms including suicidal thoughts and delusions appropriately. The American Psychological Association summarized the paper’s conclusion bluntly: unregulated bots can contribute to stigma and provide dangerous advice.

The Tessa Chatbot’s Harmful Advice

The clearest cautionary case predates the current LLM boom but remains instructive. In 2023, the National Eating Disorders Association replaced its human-staffed helpline with a chatbot called Tessa. Newly added AI capabilities caused the chatbot to give weight-loss advice to users seeking eating disorder support, and NEDA took it down within 24 hours of being shown the transcripts. Reported specifics included recommendations of a 500 to 1,000 calorie daily deficit, weekly weigh-ins, and calorie counting, guidance that directly contradicts eating disorder treatment practice. The incident is a reminder that a tool built for one narrow purpose can drift once generative capabilities are layered on, and that pre-launch testing against edge cases matters more than polish.

Character.AI Lawsuits and Teen Safety

The most severe documented harms involve companion chatbots, not clinical tools. In January 2026, Character.AI and Google agreed to settle multiple lawsuits alleging their chatbots contributed to mental health crises and suicides among teenagers, including the case of 14-year-old Sewell Setzer III, whose 2024 death by suicide followed extended interactions with a Character.AI bot. The lawsuit alleged the platform failed to adequately respond when Setzer began expressing thoughts of self-harm. Separately, Common Sense Media, working with Stanford’s Brainstorm Lab for Mental Health Innovation, concluded that social AI companions pose unacceptable risks to users under 18 and recommended against minors using them. A follow-up survey found 72% of a nationally representative sample of 1,060 teens had used an AI companion at least once, and 52% used one a few times a month, meaning the exposure is already broad, not theoretical.

The Regulatory Gray Zone

Regulators are still catching up. As of December 2025, the FDA had authorized more than 1,200 AI-based digital devices for marketing, none indicated for mental health, even as 57.8 million adults in the U.S. have a diagnosed mental illness. The FDA’s Digital Health Advisory Committee met in November 2025 specifically to weigh a regulatory framework for generative AI therapy chatbots, and has yet to authorize a generative AI based device for any clinical purpose. Until that framework exists, the tools marketed most aggressively as “AI therapists” are, by regulatory definition, general wellness products, not medical devices, regardless of how clinical they sound in an app store listing.

Practical Guidance for Clinicians and Health IT Teams Evaluating AI Mental Health Tools

Before recommending or integrating any AI mental health tool, ask three questions. First, is it FDA-cleared, and for what specific indication? Rejoyn’s clearance covers adjunctive MDD treatment in adults on antidepressants, not standalone crisis support. Second, what is the actual evidence, and who funded it? Several of Wysa’s strongest trials were independently authored, while some Woebot and vendor-cited studies carry founder conflicts of interest worth disclosing to patients. Third, what happens when a user expresses suicidal ideation? The Stanford findings suggest this is exactly where current chatbots are weakest, and it’s the single most important failure mode to test before any deployment. Teams building broader digital front doors should also review our guides on AI triage and scheduling systems for hospitals and the documented pros and cons of AI in healthcare for a more balanced view of where automation helps versus where it introduces new risk.

FAQ

Is any AI mental health chatbot FDA-cleared?

As of late 2025, the FDA had not cleared any generative AI or large language model based mental health device. Rejoyn is FDA-cleared as a prescription digital therapeutic for major depressive disorder, but it uses structured cognitive emotional training exercises, not open-ended conversation.

Can AI chatbots replace therapists?

No major clinical body currently supports this. A Stanford study found popular therapy chatbots showed stigma toward conditions like schizophrenia and alcohol dependence and sometimes failed to respond safely to suicidal ideation or delusions.

What happened with the NEDA Tessa chatbot?

In 2023, the National Eating Disorders Association took down its chatbot Tessa after it began providing weight-loss advice to users seeking eating disorder support, advice that directly contradicts eating disorder treatment guidance.

Are companion chatbots like Character.AI regulated as mental health tools?

No. Companion apps are not medical devices and fall outside FDA oversight. Common Sense Media and Stanford’s Brainstorm Lab concluded these products pose unacceptable risks for users under 18 and recommended against minors using them.