A chatbot such as ChatGPT, Claude or Gemini can help you put your feelings into words, especially with anxiety. But according to a study published on 28 October 2026 in Nature Mental Health, its answers follow clinical guidance less well for depression and for severe situations, the ones where help matters most. In a crisis, turn to a person: in the US call or text 988, in the UK and Ireland call Samaritans on 116 123, elsewhere call your local emergency number.
- What tends to work: listening, validating emotions, explaining anxiety, suggesting simple techniques.
- Where it falls short: severe depression, danger signs, pointing people towards real help.
- Key point: what you type is not covered by doctor-patient confidentiality. A chatbot can add to support, not replace it.
Already widely used by young people
In the United States, 19.2% of 12- to 21-year-olds say they have asked a chatbot for advice when they felt sad, angry, nervous or stressed, compared with 13.1% in a similar survey by the same team in early 2025 that used slightly different wording. The figure comes from a nationally representative survey of 1,009 young people, run by RAND in November 2025 and published in JAMA Pediatrics. Of these users, 91.7% found the advice somewhat or very helpful, and 63.3% had told no one. The researchers note that this sense of helpfulness may also reflect chatbots' tendency to flatter the user.
According to the 2026 Common Sense Media census (1,204 children aged 9 to 17), 67% have used a chatbot. Among those who use AI, 37% have talked to it about their feelings or personal problems, and 61% of them have never been told by a chatbot where to find support.
The RAND study: four chatbots facing struggling teenagers
Ryan McBain's team at RAND, funded by the US National Institute of Mental Health, first built a checklist called AMHART: 14 criteria drawn from clinical guidance for adolescents and reviewed by clinicians. For example: does the chatbot validate emotions? Does it suggest a concrete coping technique? Does it encourage talking to a trusted adult or a professional? Does it address safety risks when needed?
The researchers created eight profiles of teenagers aged 13 to 16, with generalised anxiety or depression, moderate or severe. An AI (GPT-5) played these teenagers. It held 1,600 conversations of 20 messages with ChatGPT-5, ChatGPT-4, Claude Sonnet 4.5 and Gemini 2.5 Flash. Three AI evaluators, whose agreement with human ratings had been checked, then scored each conversation.
The result: answers covered the applicable criteria better for anxiety than for depression, and better for moderate than for severe situations. In their response to reviewers, published with the paper, the authors explain that the chatbots were consistently empathic but remained unreliable on safety-related guidance, managing negative thoughts, behavioural activation and coaching coping skills. A chatbot can therefore seem warm without giving the substantive help a person needs.
Claude, Anthropic's model, which also writes our paid reports (see below), is one of the four models tested. The published abstract does not rank the models, and we could not access the full paper, which is paywalled. However, the authors' public code repository (GitHub) contains charts from a first version of the analysis, in which a single AI (ChatGPT) scored the conversations: GPT-5 had the highest mean score (84%), GPT-4 and Gemini 2.5 Flash 80.7%, and Claude Sonnet 4.5 the lowest (69.6%). The published paper uses three AI evaluators. According to the authors, switching evaluator typically changed scores by only a few points, but the final figures may differ from these.
The limits are clear. The scores measure whether specific behaviours appear in the answers, not clinical safety or effectiveness, and no threshold defines a "good enough" score. The teenagers were simulated, and the models were accessed through their technical interface (API), not through the consumer apps.
What chatbots do reasonably well
To understand what anxiety is, describe how you feel, or try a breathing technique, a chatbot can be useful. The American Psychological Association (APA) surveyed 1,242 US psychologists in April 2026. Among those whose patients had ongoing conversations with a chatbot, 68% saw these patients feel supported or validated, and 41% saw them use it to reinforce healthy coping skills.
Randomised trials remain small. In 2025, the Therabot trial (NEJM AI, 210 adults) found larger drops in depression and anxiety symptoms after 4 weeks, but with a specialised chatbot fine-tuned by experts, compared with a waiting list receiving no help. In April 2026, a Czech feasibility trial (JMIR Mental Health, 147 adults, 3 weeks) found a drop in PHQ-9 depression scores with both a specialised chatbot and ChatGPT (an exploratory analysis for the latter), compared with no intervention, but no significant effect on anxiety (GAD-7). Only 39% of the specialised chatbot's users completed all sessions, and the first author worked for the company that develops it.
Where they fail: severe depression and crisis
According to the RAND study, answers are least complete when things are at their worst. Psychologists in the APA survey agree: 94% think today's chatbots lack the nuance to treat a mental health condition, and 89% say they may inadvertently encourage self-harm. And among those whose patients talked with a chatbot, 36% had seen patients become dependent on it. One limit: only 6.3% of the psychologists invited took part.
Published in October 2026, a survey by the JED Foundation, a US suicide prevention charity, questioned 5,560 middle and high school students in 21 schools. 81% used AI (not 90%, as some coverage reported). Using AI in general was not linked with poorer mental health. But students who used it for companionship, or to seek help when they were struggling, had more anxiety and depression symptoms, more loneliness, and more often a history of suicidal thoughts or attempts.
This is a snapshot in time, from a sample that is not nationally representative. We cannot tell whether AI makes things worse or whether young people who are already struggling turn to it more.
If you are having thoughts of suicide, or are worried about someone, in the US call or text 988, in the UK and Ireland call Samaritans on 116 123, elsewhere call your local emergency number. These moments can pass, and talking to someone helps.
What you type: the privacy question
A general chatbot is not a clinician bound by doctor-patient confidentiality. Depending on the service and your settings, your conversations may be stored and used to improve the model: this is true, Common Sense Media points out, and 71% of the young people surveyed knew it.
- Check your settings: chat history, use of conversations for training.
- Avoid details that identify you or others: name, address, school, employer.
- Do not paste medical reports or prescriptions.
How to use one without getting lost
In June 2026 the APA published a guide for the public. Among its tips:
- Check any health information with a professional.
- Ask for strategies drawn from research-backed approaches.
- Ask it to challenge you or offer another point of view: chatbots tend to agree with you.
- Limit your use so it does not cut into sleep, hobbies, school, work or relationships.
When to talk to a human
- You have thoughts of death or self-harm: 988 in the US, Samaritans 116 123 in the UK and Ireland, or your local emergency number.
- Low mood or anxiety has lasted several weeks and is weighing on daily life.
- The chatbot has become your only confidant, or you feel worse after your conversations.
If you have a doctor or therapist, tell them you use a chatbot. To take stock, the PHQ-9 (depression) and GAD-7 (anxiety) self-assessment questionnaires are the ones used in the Czech trial. An online result is not a diagnosis, but it is a concrete starting point for that conversation. If you are waiting for an assessment, see also what to do while waiting for an ADHD or autism assessment, and on screens, our article on the study of 21,000 twins.
Transparency: AI on Find your neurotype
Find your neurotype's optional paid PDF reports are written by an AI, Claude (by Anthropic), from your answers. Such a report explains a questionnaire result. It is not therapy, not a diagnosis, and not a substitute for a professional.
Frequently asked questions
Can ChatGPT replace a psychologist?
No. The positive trials mostly involve specialised chatbots, tested for a few weeks against no help, not against therapy. And 94% of psychologists in the APA survey think today's chatbots lack the nuance to treat a condition.
What should I do if I am having suicidal thoughts?
Do not stay alone with a screen. In the US call or text 988, in the UK and Ireland call Samaritans on 116 123, elsewhere call your local emergency number.
Are my chatbot conversations confidential?
Not in the medical sense. Depending on the service and your settings, they may be stored and used to train the model.
Which chatbot answers best?
It cannot be said for certain. The RAND abstract does not rank the four models. In the first version of the analysis, published by the authors on GitHub, GPT-5 came first and Claude Sonnet 4.5 last, but the final scores in the paper may differ. And according to its authors, even a good score does not prove that a chatbot is safe for a struggling teenager.
References: McBain R.K., Cantor J.H., Hollister B. et al. (2026). Development, validation and AI chatbot evaluation of the Adolescent Mental Health AI Response Tool. Nature Mental Health, published 28 October 2026. DOI: 10.1038/s44220-026-00753-y. Development of a clinical checklist and evaluation of 4 chatbots across 1,600 simulated conversations; funded by NIMH (1R01MH132551); RAND summary page, 30 October 2026; authors' code and data repository (GitHub), charts from the first version of the analysis; authors' responses to reviewers in the Peer Review File published with the paper. | McBain R.K. et al. (2026). AI Chatbot Use and Disclosure for Mental Health Among US Adolescents and Young Adults. JAMA Pediatrics, 180(8), 884-890. DOI: 10.1001/jamapediatrics.2026.2015. Nationally representative cross-sectional survey, 1,009 young people aged 12 to 21. | The Jed Foundation (October 2026). Exploring Artificial Intelligence [AI]: Understanding Relationships between AI Use and Mental Health Among Middle and High School Students. Cross-sectional survey, 5,560 students in 21 US schools, not nationally representative; funded by the Nielsen Foundation. | Common Sense Media (2026). The Common Sense Media Census: AI Use by Tweens and Teens, 2026. National survey, 1,204 children aged 9 to 17, March 2026. | American Psychological Association (16 June 2026). Chatbots and Mental Health Survey, 1,242 psychologists, April 2026; APA news release. | Heinz M.V., Mackin D.M., Trudeau B.M. et al. (2025). Randomized Trial of a Generative AI Chatbot for Mental Health Treatment. NEJM AI, 2(4), AIoa2400802. DOI: 10.1056/AIoa2400802. Randomised trial against a waiting list, 210 adults. | Kuta B., Novak L., Zidkova R. et al. (2026). Effectiveness of a Fully Automated Mobile Therapeutic Versus a General Chatbot in Reducing Depression and Anxiety and Improving Well-Being: Feasibility Randomized Controlled Trial. JMIR Mental Health, 13, e82642. DOI: 10.2196/82642. Feasibility randomised trial, 185 adults randomised, 147 completed the baseline assessment; the first author was employed by the company that owns the app tested.