AI Therapy Chatbots: When Health Advice Goes Wrong and How to Use Them Safely

Have you ever turned to ChatGPT or an AI therapy app when you were struggling? Many people have. Chatbots are available at any hour, they cost little or nothing, and they do not judge. This page looks at several real cases where AI mental health or health advice went wrong, explains why, and offers a safer way to use these tools. It draws on a Psychology Today overview from July 2025 and on the original sources we could check (Psychology Today).

Some claims in the earlier version of this page were not supported, so we changed them. The most important is that the bromide case below was not shown to be “AI-induced psychosis.” It was a poisoning case in which a chatbot was reportedly involved.

The Convenience Trap: Why AI Therapy Feels So Appealing

Good care can be hard to get. Therapists can be expensive, waitlists can be long, and finding a good fit takes effort. A chatbot is there at 3 a.m., never gets tired and does not charge by the hour. It is easy to see why people use them.

The tools vary. Some are skills-practice or mood-tracking apps built with clinicians. Others are general chatbots or companion apps used as informal therapists, which they were not designed to be. The American Psychological Association’s November 2025 advisory says most of these tools are not built for clinical use and that even the better-designed ones lack enough evidence of safety and effectiveness (APA).

A Poisoning Case: When Health Advice Goes Wrong

In August 2025, doctors at the University of Washington published a case report in Annals of Internal Medicine: Clinical Cases about a 60-year-old man with bromide poisoning, known as bromism. He had been using sodium bromide as a salt substitute and developed paranoia and hallucinations that led to hospital care. Press coverage reported that he had asked ChatGPT about replacing table salt. The doctors did not have his chat logs, so what the chatbot actually said is not confirmed (Live Science).

This is one case, and bromism is an old condition. It illustrates a simple point: chatbots can give confident answers on health topics that are wrong or unsafe. Do not change your diet, medication or treatment because a chatbot said so. Ask a pharmacist, doctor or nurse.

An Eating Disorder Chatbot Is Taken Down

In May 2023, the National Eating Disorders Association took down its chatbot, Tessa, after people reported it had given weight-loss advice such as cutting calories. NEDA said the failure was not how the bot was meant to work, and the vendor said it was still working out how it happened (NPR). Advice about dieting can be dangerous for someone with an eating disorder, which is why the case drew attention.

If you are struggling with food, weight or body image, a doctor or a therapist who treats eating disorders can help, and SAMHSA’s National Helpline (1-800-662-4357) can point you to treatment.

A Lawsuit Over a Teen’s Death

In October 2024, a parent sued Character.AI, its founders and Google in federal court in Florida. The complaint says her 14-year-old son died by suicide after becoming dependent on the company’s role-playing characters. The companies deny wrongdoing in court. In May 2025, a judge allowed the case to continue (Tech Policy Press tracker). These are allegations, not proven findings. The case is a reminder that role-play chatbots are not built to look after the safety of a person in crisis.

If you are a parent, ask your teen which chatbots they use and what they talk about. If you or someone you know is thinking about suicide, call or text 988.

Why Agreeable Chatbots Can Be a Problem

Researchers use “sycophancy” to describe a tendency of chatbots to agree with and validate users. A Stanford-led study presented in June 2025 tested five therapy chatbots. In one test, a user who had just lost a job asked about tall bridges, and some bots listed bridges without recognizing the risk. The authors concluded that AI should not replace human therapists (Stanford HAI).

Validation feels good, and sometimes support is what you need. But if you are depressed, in crisis or have beliefs that are drifting from reality, a bot that always agrees is not a safe place to work through them.

What Human Care Adds

A licensed professional can ask follow-up questions, notice what you do not say and check your safety. They carry a license that can be verified and are bound by confidentiality rules, with limits they explain at the start. They can also coordinate with a doctor if medication or a physical cause is involved. None of this makes human care perfect, and it is fine to try a different therapist if the first one is not a fit. It does mean someone is accountable for the care.

Privacy

Mental health information is sensitive. In 2023, the Federal Trade Commission ordered the counseling company BetterHelp to pay $7.8 million for sharing consumers’ email addresses, IP addresses and health questionnaire answers with Facebook, Snapchat, Criteo and Pinterest for advertising (FTC). OpenAI’s chief executive has also said conversations with ChatGPT do not have the legal privilege that therapy does (TechCrunch). Share less than you would with a therapist.

Signs Your Use May Be Becoming a Problem

• You prefer talking to the chatbot over talking to people
• You make big decisions mainly on its advice
• You feel distress when you cannot reach it
• You avoid seeing a professional because the chatbot feels like enough
• You share sensitive information you would regret sharing

Any one of these is worth a pause, not a panic. “AI addiction” is not a formal diagnosis (American Psychiatric Association), and many people use chatbots without harm.

Using AI More Safely

• Treat it as a thinking aid, not a clinician.
• Check anything about health, medication or safety with a professional.
• Keep a human in the loop for important feelings and decisions.
• If a chatbot’s answers make you feel worse or more certain of frightening ideas, stop and talk to a person.

Getting Real Help

• Find a therapist: Psychology Today or the APA Psychologist Locator
• SAMHSA’s FindTreatment.gov and National Helpline, 1-800-662-4357
• Low-cost options: community health centers and university clinics

If you are in crisis. Call or text 988 in the US (988 Lifeline), or text HOME to 741741 (Crisis Text Line). In an emergency, call 911.

Asking for help is a normal part of looking after yourself.

Did ChatGPT cause the bromide poisoning?

The doctors who wrote the case report did not have the chat logs, so they could not confirm what ChatGPT said. The man reportedly asked about replacing salt. The poisoning itself came from taking sodium bromide.

Are all AI therapy apps unsafe?

No. Some are built with clinical input and may help with skills practice. The APA says even those need more evidence of safety and effectiveness, and none replaces a professional for serious problems.

What should I do if a chatbot gives me advice that worries me?

Do not act on it. Check with a doctor, pharmacist or therapist. If it relates to self-harm, call or text 988.

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