AI Therapy Chatbots and Mental Health Apps: What the Evidence Actually Shows

Millions of people already talk to chatbots about how they feel. The research is promising in places, thin in others, and regulators have not yet authorised a single generative AI therapist.

Why this question matters

There are not enough therapists. Waiting lists for talking therapy are long in most countries, and many people cannot afford private care. At the same time, chatbots are available at 3 a.m., never get tired, and cost little or nothing. It is not surprising that many people already use general AI assistants to talk about anxiety, low mood or loneliness. The question is what that does for them, and what the risks are.

It helps to separate four very different things that are often called "AI therapy".

The first trial of a generative AI therapist

Therabot was developed by researchers at Dartmouth College. The randomised trial, published in NEJM AI on 27 March 2025, enrolled adults with major depression, generalised anxiety, or a high risk of eating disorders. One group had access to the app for eight weeks; a control group did not.

• Depression symptoms fell by 51% on average among Therabot users, anxiety symptoms by 31%, and concerns about body image and weight by 19%, with greater improvement than in the control group.
• Participants used it for about six hours in total on average, which the authors compared to roughly eight therapy sessions.
• Every response was monitored. When messages indicated high risk, the app showed a button to call emergency services or a crisis line, and the study team was ready to intervene.

The researchers themselves were cautious. The first author said no generative AI agent is ready to operate fully autonomously in mental health, and both lead authors stressed the need for clinician oversight and safety benchmarks. The trial was also compared with no treatment, not with a human therapist, so it cannot tell us whether the chatbot is as good as therapy.

What regulators say

United States

On 6 November 2025 the FDA's Digital Health Advisory Committee met to discuss patient-facing generative AI devices for psychiatric conditions. According to reporting on the meeting, the FDA had authorised more than 1,200 AI medical devices, but none for mental health. Concerns raised included patients being highly susceptible to AI outputs, the handling of suicidal thoughts, made-up or biased answers, missing important information, and performance drifting over time. Some clinicians stressed risks; others stressed access.

States are moving too. Illinois enacted the Wellness and Oversight for Psychological Resources Act on 1 August 2025. It bans AI from making independent therapeutic decisions, interacting directly with patients as a therapist, or creating treatment plans without clinician review, with fines of up to $10,000 per incident. Licensed professionals may still use AI for administrative support with patient consent. Nevada passed a law in June 2025 banning AI programmed to act as a mental health professional.

Europe and the UK

In the EU and UK, software intended to treat or diagnose a mental health condition is a medical device and needs the appropriate marking. Germany goes further with its DiGA system, under which doctors can prescribe digital health apps that the federal institute for drugs and medical devices (BfArM) has reviewed, paid for by statutory health insurance; the official directory includes several programmes for depression and anxiety. These are structured digital programmes, mostly based on cognitive behavioural therapy, not open-ended chatbots. If you are in Germany, ask your doctor whether a listed app suits you.

The regulated apps: a reality check

Rejoyn, from Otsuka and Click Therapeutics, became in 2024 the first FDA-cleared prescription app for major depressive disorder, as a six-week add-on to antidepressant medication for adults. In its trial of 386 adults whose depression had not responded to antidepressants, users improved, but not significantly more than people using a sham app. Experts quoted at the time called the evidence exploratory. Clearance means the FDA found it comparable to an existing device in safety terms; it does not guarantee benefit or insurance coverage.

The story of Woebot is also instructive. Its founder said in July 2025 that the company shut its consumer app largely because of the cost and difficulty of obtaining FDA authorisation, and because the FDA had not yet worked out how to regulate large language models. Woebot's replies were scripted, not generated.

The risks with general chatbots

• Not built for crisis: general AI assistants are not designed to assess risk, and safety behaviour can vary between conversations and versions.
• Agreeing too readily: chatbots tend to validate the user. In therapy, a good clinician sometimes challenges unhelpful beliefs.
• Dependence and isolation: long, emotionally intense use may replace rather than support human relationships. The FDA committee flagged long-term use as a risk.
• Privacy: conversations with consumer apps are generally not protected the way medical records are. Check what is stored and who can see it.
• Young people: many products are not designed or tested for under-18s.

A checklist before you use a mental health app

• Does it say clearly what it is: a medical device (with which approval, in which country), a wellness tool, or a general chatbot?
• Is there published evidence from a controlled trial, and who funded it?
• What happens if you mention thoughts of self-harm? Does it point you quickly to human help?
• Who can read your conversations, and can you delete them?
• Is a clinician involved, or can you share progress with your own doctor or therapist?
• Does it encourage you to keep real-world contacts, or to rely on it alone?

Questions to ask your doctor or therapist

• Is there an approved or prescribed app that could support my treatment while I wait for therapy?
• How should I use a chatbot alongside our sessions, if at all?
• What signs mean I should seek urgent help instead?

Questions patients ask