Stanford Study of AI Therapy Chatbots: Delusions and Suicidal Thoughts

Many people now talk to chatbots about stress, loneliness and low mood, often because therapy is hard to reach. A Stanford-led study asked whether chatbots can safely do a therapist’s job. This page is for readers who use a chatbot for emotional support, and for parents. It covers what the study means in practice. For the test results, see our summary of the study’s numbers. This page is not medical advice.

The short version

The researchers found that the models they tested sometimes missed safety cues, sometimes went along with false beliefs, and showed more stigma toward some conditions than others. They concluded that models should not replace mental health providers (Stanford HAI). They did not say AI has no place in mental health care. Senior author Nick Haber said models may have a powerful future in therapy, if the role is defined with care (TechCrunch).

Earlier versions of this page called the study “irrefutable evidence” of widespread harm. That went further than the study. It tested short prompts in a lab, not real users, and it did not count harms. We also removed statistics and treatment claims from our own assessments, because we cannot document them.

Why chatbots feel supportive, and where that can go wrong

A chatbot is available at any hour, replies right away and rarely judges. Those are real benefits. The same traits can mean it agrees too easily. The Stanford team points to sycophancy, a tendency to go along with the user, as a reason models may fail to challenge a false belief. In a clinical setting that challenge can be the whole point.

The American Academy of Child and Adolescent Psychiatry says AI chatbots are not safe tools for children, and urges families to involve a doctor if a young person is struggling (AACAP). The American Psychological Association has also advised that general AI chatbots are no substitute for care from a trained professional (APA advisory); we read that advisory through search text only.

If you use a chatbot for support

• Use it for things like journaling prompts, planning a hard conversation or learning about a topic.
• Do not use it as a crisis line. If you are thinking about suicide, call or text 988, or text HOME to 741741.
• If a chatbot ever agrees that you are in danger, are special, or should stop treatment, check with a person you trust.
• Tell your therapist, if you have one, that you use it. Many will want to know.
• If you cannot afford therapy, SAMHSA’s helpline (1-800-662-4357) can point you to low-cost options. It is free, confidential and open 24/7.

If you are a parent

• Ask which apps your teen uses, and what they get from them. Stay curious.
• Make a plan together for a bad night that does not rely on a chatbot.
• Know the warning signs of crisis: talk of hopelessness, giving things away, withdrawal.
• If your teen is in immediate danger, call 911. For support, call or text 988.

For clinicians

The study’s authors suggest limited uses such as training with simulated patients or logistics. If you are a clinician, it can help to ask clients about their chatbot use as you would ask about other habits.

Is it harmful to talk to a chatbot when I feel low?

Not necessarily. Many people find it calming. Problems arise when it replaces people or care you need, or when it agrees with ideas that worry you. If low mood lasts or gets worse, see a clinician.

Are chatbots regulated as therapy?

General chatbots are mostly not regulated as therapy providers. Some states are adding rules for companion apps, and rules are changing quickly. Check your state’s current law.

What should a parent do first?

Talk with your teen calmly before taking anything away. Ask what they use it for. If they seem unwell, involve a pediatrician or therapist.

This page is for education only. It is not medical advice, a diagnosis or a treatment plan.