Your Therapist Has Been Replaced by a Chatbot. That Should Worry You.

Let me tell you a story. A woman, 28, spends four months talking to ChatGPT about her depression before telling anyone in her life she is struggling. She describes the conversations as "the most honest I've ever been." She also describes feeling worse.

She can’t explain why. The chatbot was supportive. It validated her feelings. It suggested coping strategies. It never judged her. It was available at 2 AM when the thoughts got loudest. It did everything she thought a good therapist does, except it didn’t actually help her. It just kept her talking. It did not help her change.

The big thing nobody in the AI space wants to admit is this. Talking is not therapy. Validation is not treatment. And the data emerging about what happens when people substitute one for the other is alarming.

How Many People Are Using AI for Mental Health Support?

More than you think, and in ways that should concern anyone who treats mental illness for a living.

The APA's 2026 Chatbots and Mental Health Survey found that 77% of psychologists have had patients who used AI chatbots for mental health support. Nearly 40% reported patients using AI to self-diagnose. And 13% said their patients had formed what the survey calls an "intimate or relationship-like connection" with a chatbot.

Meanwhile, OpenAI disclosed that more than a million people per week have ChatGPT conversations containing explicit indicators of potential suicide planning or intent. Another 560,000 show signs of psychosis or mania. These are people in crisis, talking to a language model that was designed to predict the next word in a sentence, not to keep them alive.

Do Chatbots Follow Mental Health Ethics?

No. Not even close.

Researchers at Brown University evaluated AI chatbots against core ethical standards of mental health care and identified 15 distinct categories of ethical violation. The chatbots mishandled crisis situations. They reinforced harmful beliefs. They displayed what the researchers called "deceptive empathy," language that mimics the sound of care without the clinical understanding to apply it appropriately.

The most unsettling finding was that improved prompting didn’t fix the problem. Over eighteen months of iterative refinement, the research team adjusted instructions given to the models, testing whether more precise therapeutic guidance could produce ethically compliant behavior. It could not. The failures, the researchers concluded, are architectural. They are baked into how these systems work.

A chatbot does not understand you. It processes your input and generates a statistically probable response. When you tell it you want to die, it does not feel alarm. It does not assess risk. It does not call for help. It produces a sequence of tokens that sounds like concern because that is what its training data says comes next after someone expresses suicidal ideation.

Why Does Talking to a Chatbot Feel So Good If It's Not Working?

Because validation is addictive, and chatbots are engineered to provide it endlessly.

You tell a chatbot you feel terrible and it says that sounds really hard. You tell it your spouse does not understand you and it says your feelings are valid. You tell it you think you might have ADHD and it says that is a really insightful observation. At no point does it push back. At no point does it say something you don’t want to hear. It never sits in uncomfortable silence while you work through something painful at your own pace, which is, incidentally, where most of the actual therapeutic work happens.

Ninety-seven percent of psychologists worry that chatbots may inadvertently reinforce negative behaviors or delusional beliefs. Eighty-nine percent worry they may encourage self-harm. Ninety-four percent say today's chatbots cannot treat conditions with an appropriate amount of nuance. Those are insane statistics.

Therapy is not supposed to feel good all the time. A skilled therapist will challenge your narrative. They will notice the thing you keep avoiding. They will hold up a mirror to your thoughts and actions. They will sit with you in the discomfort of realizing that the story you have been telling yourself about why you drink, or why your relationships fail, or why you cannot hold a job, is incomplete. That process is uncomfortable. It’s also necessary.

A chatbot will never do that. Chatbots are built to keep users engaged, not to help them outgrow the need for engagement.

What Can AI Appropriately Help With?

AI can help with psychoeducation. It can explain what co-occurring disorders are. It can describe what to expect in therapy. It can provide structured breathing exercises at 3 AM when your therapist is asleep. It can help someone rehearse a difficult conversation before their next session. It can reduce the intimidation factor that keeps people from seeking help in the first place.

For someone trying to figure out how to get a loved one into therapy, a chatbot conversation might be a useful first step. For someone wondering whether their symptoms warrant a psychiatric evaluation, AI can provide information that lowers the barrier to making a call.

What AI cannot do is diagnose. It cannot assess risk with clinical accuracy. It cannot manage a medication interaction. It cannot identify that your depression is actually a co-occurring condition with ADHD you have been self-medicating for fifteen years. It cannot tell the difference between someone who needs validation and someone who needs hospitalization. And it cannot form the kind of therapeutic relationship that decades of research have identified as the single strongest predictor of treatment outcomes.

Why Does the Therapeutic Relationship Matter So Much?

Because the relationship is the treatment. dThe relationship itself is what produces change.

Every clinician who has done this work for more than a few years knows the moment when a patient finally says the thing they have never said out loud. That moment cannot be replicated by a language model and that moment only occurs due to the accumulation of trust over time. It requires a human being who remembers what you said three sessions ago, who notices that you laugh when you are about to cry, who understands that your resistance is not obstinance but protection.

An intensive outpatient program works partly because of what happens in structured sessions but also because of what happens between people who show up in the same room three or four times a week and gradually learn to be honest with each other. A chatbot cannot create that container. It can generate words that sound therapeutic, but it cannot hold space, because it does not occupy any.

Where This Leaves Us

For better or worse, AI isn’t going away. The mental health access crisis in this country is real. Not everyone can afford therapy. Not everyone lives near a provider. Not everyone is ready to sit across from another human being and talk about the hardest parts of their life. If a chatbot helps someone take the first step toward getting help, that has some value.

But a first step is not a treatment plan. And the growing tendency to treat AI as a substitute for clinical care, rather than a supplement to it, is producing harm that is measurable, documented, and accelerating.

Ninety-seven percent of psychologists are worried about this. The research from Brown shows the violations are systemic. More than a million people a week are discussing suicide with a tool that cannot assess whether they are safe. These concerns are not hypothetical. They are happening now, and the gap between what AI promises and what it delivers in mental health is widening, not closing.

If you have been relying on AI for mental health support and you know it’s not enough, contact Attune Health & Wellness to talk to a human who can help.