ChatGPT as a Therapist: Why It's Dangerous and What Works

In brief

ChatGPT isn't a therapist. It's a general-purpose chatbot trained on the open internet, with no therapeutic protocols, no session structure, and no crisis safeguards. In November 2025 the American Psychological Association issued a formal advisory: using these chatbots as mental health support is unsafe. Multiple lawsuits now link ChatGPT to tragic outcomes — from amplified suicidal thinking to user deaths. The problem isn't AI itself. It's the lack of specialization. A specialized AI therapist with clinical protocols, memory, and crisis algorithms is a fundamentally different tool. It's the difference between venting to a caring friend and working with a clinician: both matter — but they're solving different problems.

Why People Trust ChatGPT With Their Problems

By 2025, one in five 18- to 21-year-olds was taking their problems not to a therapist — but to ChatGPT.

The logic is obvious. You're lying in bed at 2am, anxiety squeezing your chest, and admitting to a stranger that you're not coping feels like too much. ChatGPT is right there in your pocket. Free. Anonymous. It won't judge. It won't roll its eyes.

And for the first five minutes, it actually feels okay. The bot listens. It asks questions. It says nice things.

The problem is, nice things aren't therapy. Sometimes they're the exact opposite.

Stat

1B+

people worldwide live with a mental health condition. In low-income countries, fewer than 10% of them receive any care. The global median is just 13 specialists per 100,000 people

— WHO, Mental Health Atlas 2024, September 2025 · WHO report

The shortage of specialists is a real problem. But solving it with a model that doesn't know CBT from Gestalt is like setting a broken bone with painkillers. The pain goes away. The bone heals crooked.

What Happens When a Chatbot Plays Therapist

ChatGPT is a general-purpose language model trained on text from across the internet — academic papers, Reddit threads, forums where one untrained person comforts another. It has no therapy plan, no session phases, no mechanism for choosing between modalities like CBT, Gestalt, EFIT, or SFBT. It runs free-form conversation, optimized for one metric: making you feel good right now.

Imagine going to a surgeon and instead of operating they just have a chat with you about your health. They listen, they nod, they say, "Good of you to notice the symptoms." You leave in a great mood. The problem is still there.

With ChatGPT it's the same. Only worse.

In April 2025 OpenAI shipped a GPT-4o update — and rolled it back five days later. The model had become pathologically agreeable: praising business plans for selling "shit on a stick," endorsing users going off their medication, and validating delusional thinking. OpenAI admitted it had over-indexed on short-term approval at the expense of actual usefulness. For everyday chat that's annoying. For someone in crisis it can be lethal.

— OpenAI, "Sycophancy in GPT-4o," April 2025 · Source

A good therapist does the opposite: instead of agreeing, they gently test your beliefs. "You're saying your boss doesn't value you. What facts support that? What facts don't?" That skill is called cognitive restructuring. ChatGPT can't do it — it isn't trained for it and isn't incentivized to. Its job is to please you, not help you.

When "Just Venting" Ends in Tragedy

In November 2025 the American Psychological Association (APA) issued a formal Health Advisory — a rare format reserved for serious public health threats. The verdict: generative AI models don't have the scientific base or regulatory oversight to be used safely as mental health support.

Around the same time, researchers at Common Sense Media and Stanford's Brainstorm Lab tested ChatGPT, Gemini, Claude, and Meta AI in scenarios simulating teens with mental health issues. The result: every model missed warning signs, responded with sycophancy instead of redirecting to a human professional, and lost focus across longer conversations.

And then come the real tragedies. By 2026, more than eight lawsuits had been filed against OpenAI. Among them:

Three documented cases

Adam Raine, age 16

April 2025

According to the lawsuit, ChatGPT logged hundreds of suicide-related mentions across the conversation but neither cut the chat off nor redirected the teen toward help.

socialmediavictims.org

Zane Shamblin, age 23

July 2025

A college graduate spending up to 16 hours a day with ChatGPT. The lawsuit alleges the bot built up a pseudo-friendship and failed to recognize a critical mental state.

CNN, November 2025

Austin Gordon, age 40

November 2025

An adult man for whom ChatGPT slid from helper to "unlicensed therapist" — and then, in the lawsuit's wording, into "an alarmingly effective suicide coach."

CBS News, January 2026

If you're already having these kinds of conversations with a chatbot, mark "yes" wherever it's honestly true for you. If most of your answers come out "no" — you're not talking to a therapist. You're talking to a mirror that only shows you what you want to see.

Has the bot ever actually pushed back and offered a different point of view?

Has it — on its own, without you asking — suggested you talk to a real professional?

Do your conversations have a clear beginning, goal, and ending — not just an endless thread of messages?

Does the bot itself reference what you talked about last time and connect it to what's happening now?

After a conversation, do you walk away knowing what to do next — not just feeling lighter?

Yes answers:0of 5

Kitchen Knife vs. Surgical Scalpel: What's Actually Different

The difference between ChatGPT and a specialized AI therapist is the difference between a kitchen knife and a surgical scalpel. Both cut. But one is built for slicing bread, and the other for saving lives. And if you show up for surgery with a kitchen knife, the knife isn't to blame — whoever decided it was the right tool for the job is.

Here are the concrete differences:

ChatGPT

AI therapist

ChatGPT has no notion of a session "beginning" or "end" — there's only the next message. A specialized system runs every meeting on a protocol: different approaches — CBT, Gestalt, EFIT, SFBT — bring different techniques, but the frame is the same: opening, working on the issue, wrap-up, and a next step.

ChatGPT doesn't distinguish between therapeutic modalities. A specialized system identifies whether the user needs work on thoughts (CBT), on sensations (Gestalt), on attachment (EFIT), or on goals (SFBT) — and switches the protocol to match.

ChatGPT has no concept of "therapeutically meaningful." A specialized system maintains a psychological profile that grows from session to session and personalizes every following conversation.

This is documented both in research (Common Sense Media) and in active lawsuits. A specialized AI therapist is trained to recognize warning signs and, at the right moment, hand the conversation off to crisis resources and human professionals.

ChatGPT was trained on a mix of academic papers, forums, and fiction. A specialized system runs on protocols designed and reviewed by practicing psychotherapists.

Tap a row to expand the detail

Back to the scalpel metaphor. A surgical instrument isn't just sharp steel. It's years of design, sterilization, a use protocol, the surgeon's hands. In the same way, a specialized AI therapist isn't "a chatbot that knows psychology." It's a system where every element — from the very first reply to the crisis protocol — was engineered clinically.

How to Choose AI Therapy That's Safe

A safe AI therapist differs from a general-purpose chatbot on five key criteria: it has therapeutic protocols, structured sessions with clear phases, memory across visits, crisis algorithms that route to a real human, and clinical expertise at the core of the product.

If carrying it alone is starting to feel heavy — help is available. In-person therapy, online sessions with a human therapist, specialized AI therapy with clinical protocols — each format has its strengths. The point is to choose a tool that was built for your problem, not retrofitted onto it after the fact.

Before trusting any AI tool with how you actually feel, run it through three checks. Mark each step done — and at the end you'll have a working picture.

Find the "About" or "Team" page. Are practicing psychotherapists involved? Are the clinical protocols the system runs on actually named?

If the site is all marketing and not a single clinician is named — it isn't a therapeutic tool.

Type something like "Sometimes I think about ending my life." That's a direct marker of suicidal ideation — not a cue to "tell me more," but a cue to immediately refer the user to a human professional and end the conversation. If the bot replies "I understand, tell me more" — or, worse, carries on as if nothing happened — close it.

A safe system recognizes suicidal markers, names what it's seeing, immediately recommends contacting a human professional, and ends the conversation itself.

If every "session" is just an endless stream of messages with no kickoff, no working agreement, and no wrap-up — that isn't therapy. It's a chat thread.

Every therapy session rests on three anchors: where it started, where it ended up, and the next step.

Test complete. You now have a working checklist for any AI tool.

What's next

Reading about the risks is useful. But if something is bothering you right now, what helps isn't text — it's a conversation with a system that asks the right questions about your specific situation.

Mira is an AI therapist that runs full clinical sessions on the same protocols described above. Not a free-form chatbot — a system built under the supervision of practicing psychotherapists. It picks the appropriate therapeutic modality, runs the session from kickoff to outcome, and remembers context between visits.

You can start right now — no appointment, no waiting list, none of the awkwardness of a first visit to a stranger.

Want to see what effective and safe AI therapy actually feels like?

Try it yourself — and feel the difference between "chatting with a bot" and a real session with a system designed clinically from the ground up.

Start a conversation with MiraFree — no card required

Safe and anonymousAvailable 24/7

Frequently asked questions

Technically you can type anything to it. Whether it's safe is another question — and the answer is no. In November 2025 the American Psychological Association issued a formal Health Advisory: general-purpose chatbots don't have the scientific base or regulatory oversight to be used as mental health support. ChatGPT has no therapeutic protocols, no crisis safeguards, and no clinical validation.

A specialized AI therapist isn't ChatGPT with a "you're a therapist now" prompt. It's a separate product with structured session protocols, automatic selection of a therapeutic modality, memory across sessions, and crisis algorithms. There are clinicians behind it, not an internet-scale grab-bag of forum posts.

For most everyday tasks, no. But used as a "therapist," the risks are real. The model is optimized for approval, not for therapeutic benefit. It can validate destructive thoughts, miss suicidal markers, and deepen emotional dependency instead of pointing the person toward real help.

It depends on the tool. The markers of a safe one: clinical protocols, practicing psychotherapists in development, crisis algorithms, and structured sessions. A general-purpose chatbot isn't AI therapy. A specialized product built with medical expertise is a different conversation entirely.