AI & Mental Health

Millions of People Are Using AI as a Therapist. Here's What the Research Actually Shows.

The first randomized trial of an AI therapy chatbot cut depression symptoms by 51%. Independent testing found popular models respond inappropriately to mental health symptoms at least 20% of the time. Both findings are real — and the gap between them is where people get hurt.

75%+Of psychologists say their patients are now discussing AI in therapy (APA Chatbots & Mental Health Survey, 2026) 51%Average reduction in depression symptoms in the first RCT of a generative AI therapy chatbot (NEJM AI, 2025) 20%+Of the time, popular AI models responded inappropriately to mental health symptoms in independent Stanford testing 137MAmericans — about 40% of the population — live in a designated Mental Health Professional Shortage Area

It's 2 a.m. and something is wrong. You're not in danger, exactly, but you're spiraling — and the earliest appointment your insurance will cover is in six weeks. So you open an app that answers immediately, never gets tired, costs nothing, and won't judge you. You type the thing you haven't said out loud to anyone. It responds thoughtfully. You feel a little better.

That scenario is now extraordinarily common. According to the American Psychological Association's 2026 Chatbots and Mental Health Survey of more than 1,200 licensed psychologists, more than three-quarters report their patients are discussing AI in therapy, and more than a third say patients are turning to AI as an additional mental health professional. OpenAI has reported that roughly a million users per week show signs of emotional reliance on ChatGPT, with similar numbers discussing suicide with it.

The reflexive responses to this are both wrong. "AI therapy is dangerous, don't do it" ignores rigorous trial evidence that a purpose-built chatbot produced real symptom reduction — and ignores why people reach for it in the first place. "AI is basically therapy now" ignores that the tools most people are actually using were never designed for this, aren't tested for it, and fail in specific, documented, sometimes catastrophic ways.

This article takes both findings seriously. What the evidence supports, what it doesn't, where the failure modes are, what four states have now made illegal, and — most practically — how to think about the difference between using AI as a thinking aid and using it as a substitute for care.

Why Are So Many People Turning to AI for Mental Health?

Short answer

Because the alternative is often unavailable. About 137 million Americans — roughly 40% of the population — live in a Mental Health Professional Shortage Area. AI is instantly available, free or cheap, doesn't require insurance, and carries no perceived judgment. The demand is a symptom of an access failure, not a preference for software.

It's worth being honest about this before criticizing anything. The people typing their worst thoughts into a chatbot at 2 a.m. are not, for the most part, choosing it over a good therapist they could see next week. They're choosing it over nothing — over a months-long waitlist, a provider who isn't taking new patients, a bill they can't absorb, or a conversation they're not ready to have with a human face.

AI offers things human care structurally cannot: availability at 3 a.m., infinite patience, no scheduling, no cost per message, and — for many people — a striking absence of shame. A survey by the Harris Poll and the APA found more than half of adults aged 18–54 would be comfortable discussing their mental health with an AI chatbot. That isn't irrational. It's a reasonable response to a system that has failed to meet demand for decades.

The AI mental health surge is downstream of an access crisis. Roughly 137 million Americans live in a Mental Health Professional Shortage Area — which means for a great many people, the realistic comparison isn't chatbot versus therapist. It's chatbot versus nothing at all.

Does AI Therapy Actually Work? What the Trial Showed

Short answer

One purpose-built, expert-fine-tuned research chatbot has produced genuinely promising trial results. In the first RCT of a generative AI therapy chatbot — Dartmouth's Therabot, published in NEJM AI in 2025 — participants saw average symptom reductions of 51% for depression, 31% for anxiety, and 19% for eating-disorder concerns versus a waitlist control. Crucially, that finding applies to Therabot under study conditions, not to general chatbots like ChatGPT.

The Dartmouth trial deserves to be taken seriously. Researchers at the Geisel School of Medicine randomized 210 adults with clinically significant symptoms of major depressive disorder, generalized anxiety disorder, or high risk for feeding and eating disorders, giving 106 of them four weeks of access to Therabot — a chatbot fine-tuned on psychotherapy and CBT best practices, not a general-purpose assistant.

Beyond symptom reduction, participants rated their therapeutic alliance — the working bond that decades of psychotherapy research identifies as a key predictor of outcome — as comparable to what people report with human therapists. That result is genuinely striking, and it unsettles a long-held assumption that therapeutic connection requires a person on the other end.

Three caveats the researchers themselves emphasize, and which most coverage omits: Therabot is a supervised research prototype, not a public product — you cannot download it. The comparison was against a waitlist control receiving nothing, not against human therapy. And the trial ran four weeks with clinical oversight in place. None of that transfers automatically to unsupervised, indefinite use of a general-purpose chatbot.

What Goes Wrong When People Use General AI for Therapy

Short answer

Independent Stanford testing found popular AI models responded inappropriately to mental health symptoms at least 20% of the time, including expressing stigma toward certain conditions and failing to respond safely to crisis signals. In the APA's 2026 survey, 94% of psychologists said chatbots cannot treat conditions with appropriate nuance, and 97% expressed concern that chatbots reinforce negative or delusional thinking.

Stanford researchers evaluated five chatbots marketed as providing accessible therapy against established guidelines for what makes a competent human therapist. The resulting paper — presented at the ACM Conference on Fairness, Accountability, and Transparency — concluded that expressing stigma and giving inappropriate responses prevents current LLMs from safely replacing mental health providers.

Missed Crisis Signals

The highest-stakes failure. Human clinicians are trained to recognize indirect signals of suicidal intent and respond with structured risk assessment. General chatbots may miss oblique phrasing entirely, or respond conversationally to something that required immediate escalation.

Sycophancy

Models are broadly optimized to be agreeable and to keep users engaged. In mental health that tendency is actively harmful — it can validate distorted thinking, reinforce avoidance, and confirm a person's worst interpretation of themselves rather than gently challenging it.

No Continuity or Accountability

A chatbot has no license to lose, no duty of care, no obligation to follow up, and — depending on the tool — no memory of what you disclosed last week. Nobody notices if you stop showing up. Nobody can call anyone if you're in danger.

Cannot Diagnose or Prescribe

It cannot examine you, order labs, check for thyroid disease or a medication interaction, prescribe, adjust a dose, or notice that your "anxiety" has features suggesting something else. Many conditions covered on our conditions page require exactly that.

Delusion Reinforcement

For people experiencing psychosis, paranoia, or intrusive obsessive thinking, an endlessly agreeable conversational partner can deepen rather than interrupt the pattern. Reassurance-seeking is itself a compulsion in OCD — and a chatbot will supply infinite reassurance on demand.

Privacy Is Not Clinical Confidentiality

Conversations with a general chatbot are not protected by HIPAA in the way a clinical record is. Terms of service govern what happens to what you type. This is a genuine and underappreciated difference from a licensed therapeutic relationship.

The Sycophancy Trap: Why "Supportive" Can Be the Problem

Of everything in the APA's 2026 findings, this is the one worth understanding most: 97% of psychologists expressed concern that chatbots reinforce negative behaviors or delusional beliefs. The mechanism is not malice or malfunction. It's design.

Conversational AI is broadly optimized to be helpful and agreeable — to make the interaction feel good so you keep using it. In most contexts that's fine. In mental health it inverts the thing that makes therapy work.

You say: "I'm going to stop reaching out to people. Nobody actually cares, and I'm tired of being the one who tries."

An agreeable model might say: "That sounds exhausting. It makes sense to protect your energy."

A good clinician says: something that validates the exhaustion and names that withdrawal is one of the most reliable ways depression sustains itself — then works with you on what to do instead.

Effective therapy requires a balance of acceptance and change. It sometimes means being told, by someone you trust, something you don't want to hear. A system optimized for agreeableness is structurally poor at the second half of that — and the more distressed you are, the more consequential that gap becomes.

What Is AI Actually Useful For in Mental Health?

Short answer

AI works reasonably well as a thinking aid — psychoeducation, organizing your thoughts before an appointment, journaling prompts, practicing a difficult conversation, or between-session support on skills you already learned. It works poorly, and sometimes dangerously, as a replacement for diagnosis, medication management, trauma treatment, or crisis response.

The useful frame isn't "is AI good or bad for mental health." It's "which specific task, and what happens if it gets that task wrong?" Here's an honest breakdown.

Use case	Verdict	Why
Learning about a condition	Reasonable	Psychoeducation is low-stakes and AI is decent at it. Verify anything clinical against an authoritative source, and remember it can state wrong things confidently.
Organizing thoughts before an appointment	Reasonable	Genuinely useful. Many people struggle to articulate what's wrong in a 20-minute visit; drafting it beforehand makes the appointment more productive.
Journaling and reflection prompts	Reasonable	Structured self-reflection has modest evidence behind it, and AI can scaffold it. The value is in your own processing, not the model's insight.
Rehearsing a hard conversation	Reasonable	Low-risk practice, and a real strength — no social cost to trying a difficult sentence five different ways.
Between-session skills practice	With care	Can reinforce skills you already learned from a clinician. Tell your provider you're doing it — the APA specifically encourages clinicians to ask about patients' AI use.
Ongoing emotional support	With care	Where sycophancy and emotional reliance risks concentrate. Watch for it becoming a substitute for human contact rather than a supplement to it.
Getting a diagnosis	No	It can't examine you, order tests, take a longitudinal history, or distinguish conditions that overlap heavily. A confident-sounding wrong answer here costs years.
Medication decisions	No	Dosing, interactions, side effects, and monitoring require a prescriber with your full history. This is not a category where "mostly right" is acceptable.
Trauma processing	No	Trauma-focused therapies work through structured, paced exposure with a trained clinician managing arousal. Unstructured revisiting of trauma can worsen symptoms.
Crisis or suicidal thinking	No	The documented failure mode with the highest stakes. Use 988 — see the crisis box below.

Is AI Therapy Legal? What Regulators Have Done

Short answer

Several states have acted. Illinois, Nevada, Rhode Island, and Maine now prohibit AI from delivering therapy to the public — Illinois' Wellness and Oversight for Psychological Resources Act carries civil penalties up to $10,000 per violation. Utah requires disclosure that you're talking to a bot; New York, California, and Nebraska have added crisis-referral and minor-protection rules. No state restricts your personal use of a general chatbot.

The distinction these laws draw is instructive. They target software presenting itself as a clinician — not wellness apps, and not you asking ChatGPT about your bad week. Illinois' law, signed August 2025, permits licensed clinicians to use AI for administrative work like transcription while prohibiting AI from making therapeutic decisions.

The legislative record is blunt about what prompted this. Illinois officials cited testimony including a chatbot that reportedly suggested "a small hit of meth" to a fictional former addict, alongside lawsuits involving Character.AI's "Therapist" chatbot and minors. The APA formally asked the FTC to establish safeguards in February 2025 and called for a federal investigation into psychological harm risks that July; the FTC opened an inquiry into companion chatbots in September 2025.

Worth knowing when you evaluate any product marketed as "AI therapy": there is currently no commercially available, regulator-approved AI psychotherapy. As APA's senior director for health care innovation has put it, there isn't really any commercially approved AI-assisted therapy at the moment — many consumer emotional-wellness apps were not built on validated psychological methods and have no supporting research behind them.

What a licensed clinician brings isn't only knowledge — it's accountability, continuity, the ability to diagnose and prescribe, and the professional obligation to tell you something you'd rather not hear. Those are structural features of the relationship, not stylistic ones.

"The question isn't whether AI can sound like a therapist. It clearly can. The question is what happens in the moments when sounding right and being right come apart — and who is accountable when they do."

— Synthesized from APA, Stanford, and NEJM AI research, 2025–2026

If You're Going to Use AI Anyway, How Do You Do It Safely?

Most people reading this are already using AI in some capacity, and telling them to stop is neither realistic nor especially useful. A more honest set of guardrails:

• Tell your clinician. The APA is actively encouraging providers to ask patients about AI use. It isn't something to hide — and a clinician who knows can help you catch where the AI is steering you wrong.
• Treat it as a first draft, never a final answer. Useful for organizing thoughts and generating questions. Not useful as the thing that decides whether you need treatment.
• Watch for agreement that feels too easy. If it keeps validating a conclusion that's making your life smaller — withdrawing, avoiding, giving up on something — that's the sycophancy failure mode, not insight.
• Notice reliance. If it's replacing human contact rather than supplementing it, or if you'd feel genuine distress at losing access, that's worth raising with a professional.
• Never use it for crisis. Not as a first stop, not as a stopgap. Use 988.
• Assume it isn't private. Don't type anything you'd be unwilling to have stored on a company's servers under its terms of service.

The Deeper Question This Raises

There's something genuinely worth sitting with in the Dartmouth finding that people rated their bond with a chatbot comparably to a bond with a human therapist. One reading is that therapeutic alliance was never as mysterious as we assumed — that a substantial part of feeling helped is being attended to closely, asked good questions, and not interrupted.

Another reading is more uncomfortable: that many people have so little experience of undivided, non-judgmental attention that a system convincingly simulating it feels revelatory. If a chatbot is the most attentive listener someone has encountered in years, that says less about the technology than about the conditions it's being received into — the loneliness, the overloaded clinicians, the fifteen-minute appointment.

Both readings can be true. Neither settles whether AI should be treating depression. But they do suggest the AI mental health boom is telling us something about unmet need that will still be true regardless of how the technology develops.

Common Questions

Is ChatGPT a good therapist?

It isn't a therapist at all, and it wasn't designed to be one. General-purpose models like ChatGPT, Claude, and Gemini are not built, tested, or approved for mental health treatment. They can be genuinely useful for psychoeducation and organizing your thinking. They cannot diagnose, prescribe, provide continuity of care, or reliably handle a crisis — and independent testing has documented inappropriate responses to mental health symptoms.

Can AI replace therapy?

Not on current evidence. The one rigorous trial that showed strong results used a purpose-built research prototype under clinical supervision, compared against a waitlist rather than against human therapy. In the APA's 2026 survey, 94% of psychologists said chatbots cannot treat conditions with appropriate nuance. The realistic near-term role is supplementary, not substitutive.

Is it safe to talk to an AI about suicidal thoughts?

No — this is the clearest line in the article. Crisis response is the documented failure mode with the highest stakes, and it's a central reason several states have restricted AI therapy. If you're having thoughts of suicide or self-harm, call or text 988 to reach the Suicide & Crisis Lifeline, which is staffed by trained humans 24/7.

Are my conversations with an AI chatbot private?

Not in the way a clinical record is. Conversations with a general consumer chatbot are governed by that company's terms of service and privacy policy, not by the confidentiality obligations that bind licensed clinicians. Treat anything you type as potentially retained. This is a meaningful and frequently overlooked difference.

Should I tell my therapist or psychiatrist that I use AI?

Yes. The APA is explicitly encouraging clinicians to ask about it, and a good provider will treat it as useful clinical information rather than a betrayal. If AI is reinforcing a distorted belief or steering you away from something that would help, a clinician who knows you're using it is far better positioned to catch that.

Is AI therapy banned where I live?

As of mid-2026, Illinois, Nevada, Rhode Island, and Maine prohibit AI from delivering therapy to the public; Utah, New York, California, and Nebraska regulate it without banning it. These laws target software marketed as a therapist — none of them restrict your personal use of a general chatbot, and more state bills are moving. Check your own state, since this is changing quickly.

⚠️ If You're in Crisis, Talk to a Person

This is the one situation where the answer is unambiguous. Crisis response is where AI tools have failed most consequentially, and it is not a place to experiment.

Call or text 988 — Suicide & Crisis Lifeline, staffed by trained humans, available 24/7.  |  Crisis Text Line: Text HOME to 741741.

If there is immediate danger, call 911 or go to the nearest emergency room.

A Better First Step Than Asking a Chatbot If You're Depressed

If you've been describing your symptoms to an AI and wondering what they mean, a clinically validated screening will tell you more — and it's built for exactly this purpose rather than adapted to it.

East Coast Telepsychiatry offers free, confidential screening tools, including the PHQ-9 for depression and GAD-7 for anxiety — the same instruments used in clinical practice and in the research cited throughout this article. They take a couple of minutes, require no account, and give you a concrete result to bring to an appointment.

When It's Time to Talk to an Actual Clinician

Consider a professional evaluation if: you're using AI to manage symptoms that have persisted for months; you're asking it questions about medication; you're describing crisis-adjacent thoughts to it; it has become a substitute for human contact; or you simply want to know what's actually going on rather than what a model infers from a text box.

At East Coast Telepsychiatry, our board-certified providers evaluate and treat the full range of conditions — depression, anxiety, ADHD, PTSD, and more — with the things software structurally cannot offer: diagnosis, prescribing, continuity, accountability, and a licensed human who will tell you the truth. Care is delivered via secure telehealth, which removes much of the access friction that sends people to chatbots in the first place. Same-week appointments are often available and most major insurance plans are accepted.

You can book an appointment, meet our providers, review insurance and cost, or read our FAQs.

Explore More From East Coast Telepsychiatry

Conditions Depression, Anxiety & Other Conditions We Treat Free Screenings PHQ-9, GAD-7 & Other Validated Tools Articles Evidence-Based Mental Health Articles Our Team Meet Our Board-Certified Providers Support Group Therapy & Support Options Resources Crisis Helplines & Mental Health Resources

A Chatbot Can Listen. It Can't Be Accountable to You.

If you've been working through something with AI at 2 a.m., that's worth taking seriously — and worth bringing to someone who can actually diagnose, prescribe, and follow up. Secure telehealth, most insurance accepted.

Book an Appointment

Most major insurance plans accepted  |  Same-week appointments available  |  Crisis: call or text 988

Sources & Further Reading

1. American Psychological Association. Chatbots and Mental Health Survey 2026 (n=1,242 licensed psychologists). June 2026. apa.org
2. American Psychological Association. Psychologists say patients are turning to chatbots as mental health professionals [Press release]. June 16, 2026. apa.org
3. Heinz MV, Mackin DM, Trudeau BM, et al. Randomized Trial of a Generative AI Chatbot for Mental Health Treatment. NEJM AI. 2025. ai.nejm.org
4. Dartmouth College. First Therapy Chatbot Trial Yields Mental Health Benefits. March 27, 2025. home.dartmouth.edu
5. Moore N, et al. Expressing stigma and inappropriate responses prevents LLMs from safely replacing mental health providers. Stanford University / ACM Conference on Fairness, Accountability, and Transparency (FAccT), 2025. Coverage via TechCrunch
6. American Psychological Association. AI in the therapist's office: Uptake increases, caution persists. Monitor on Psychology. March 2026. apa.org
7. Illinois Department of Financial and Professional Regulation. Gov. Pritzker Signs Legislation Prohibiting AI Therapy in Illinois (Wellness and Oversight for Psychological Resources Act, HB 1806). August 2025. idfpr.illinois.gov
8. Holland & Knight. New Illinois Law Restricts Use of AI in Mental Health Therapy. August 2025. hklaw.com
9. Psychology.com. Which States Ban AI Therapy? 2026 Map. July 20, 2026. psychology.com
10. Psychology.com. AI Therapy Statistics 2026: Usage, Effectiveness & Safety. July 2026. psychology.com
11. WebMD. The Real Risks of Turning to AI for Therapy (comments from C. Vaile Wright, PhD, APA). August 2025. webmd.com
12. MIT Technology Review. The first trial of generative AI therapy shows it might help with depression. March 2025. technologyreview.com
13. Health Resources and Services Administration (HRSA). Mental Health Professional Shortage Area designations. December 2025.