Can AI Replace Your Therapist?
The Honest Answer in 2026.
The first randomized controlled trial of an AI therapy chatbot found real, clinically significant improvements. A separate audit of the same class of tools found they routinely violate core mental health ethics — and dozens of families are now suing the makers of general-purpose chatbots over deaths linked to their use. Both realities are true at once. Here’s what that means for you.
What’s in this article
- Who’s actually using AI for mental health — and why
- The Dartmouth trial: what the best evidence actually says
- Where AI therapy fails — and how seriously
- The legal reckoning: what’s changed since spring 2026
- The gap that makes this complicated
- Therabot vs. ChatGPT vs. Wysa vs. a human therapist
- If you’re going to use an AI chatbot for support: how to do it less dangerously
- For specific audiences
- Glossary
- Common questions
First: Why People Are Actually Doing This
It’s 2am. Anxiety’s spiking. Your next therapy appointment is in ten days. You can’t afford to call your therapist out of hours. Your friends are asleep. And there’s a chatbot that will respond immediately, doesn’t judge you, and is free.
That’s why 1 in 8 US adolescents and young adults — roughly 5.4 million people — are already using AI chatbots for mental health advice, according to a RAND-led study published in JAMA Network Open in November 2025, the first nationally representative survey on this question. n=1,058, ages 12–21, surveyed Feb–March 2025 via random sampling from US households; DOI: 10.1001/jamanetworkopen.2025.42281
Among 18–21 year olds specifically, that number climbs to 22%. Two-thirds of users engage at least monthly. And 93% report finding the advice helpful.
That last number is important — and genuinely difficult. People say it helps. But subjective helpfulness and clinical safety are different things. And the same research infrastructure that produced the RAND usage data has also produced findings that should slow anyone down before treating AI as a substitute for care.
The treatment gap that’s driving this is real. For every available mental health provider in the US, there are approximately 1,600 patients with depression or anxiety alone, according to the Dartmouth researchers who ran the only RCT of a generative AI therapy chatbot. Dartmouth/NEJM AI, 2025 That’s not a minor access problem. That’s a structural failure of the care system that no amount of concern about AI will fix on its own.
“There is no replacement for in-person care, but there are nowhere near enough providers to go around.”
Nicholas Jacobson, Associate Professor of Biomedical Data Science & Psychiatry, Dartmouth — NEJM AI, March 2025
What the Best Evidence Actually Shows: The Dartmouth Therabot Trial
In March 2025, Dartmouth published what is — as of this writing — the first randomized controlled trial of a generative AI therapy chatbot. The results were genuinely striking. And the researchers’ caveats were equally striking.
Verified Research — Therabot RCT, NEJM AI, March 2025
Dartmouth’s Therabot trial: 106 participants, 8 weeks, published in NEJM AI
Participants were adults diagnosed with major depressive disorder, generalized anxiety disorder, or eating disorders. They were randomly assigned to Therabot (a purpose-built generative AI chatbot, in development since 2019 with continuous clinical oversight) or a control condition with no access.
After 8 weeks — equivalent to roughly 6 hours of interaction, or about 8 therapy sessions — the results: 51% average reduction in depression symptoms in the depression group. 31% reduction in anxiety symptoms in the anxiety group, with many shifting from moderate to mild or below clinical threshold. 19% reduction in eating disorder concerns about body image and weight, significantly outpacing the control group.
Participants also reported therapeutic alliance — the degree of trust and collaboration with a provider — at levels comparable to in-person therapy. The lead researchers described the results as comparable to gold-standard outpatient cognitive behavioral therapy.
Critical caveat from the researchers themselves: “No generative AI agent is ready to operate fully autonomously in mental health where there is a very wide range of high-risk scenarios it might encounter.” — Dr. Michael Heinz, first author, attending psychiatrist at Dartmouth Hitchcock Medical Center.
Three things about this trial matter for how you interpret the results.
One: Therabot was developed over six years with continuous input from psychiatrists and psychologists. It is not ChatGPT with a mental health prompt. Most AI tools people are actually using — including the chatbots in the RAND usage survey — are general LLMs, not purpose-built clinical tools with that level of oversight.
Two: The trial had a study team actively monitoring conversations for safety concerns in real time and was prepared to intervene. That’s not what happens when someone opens ChatGPT alone at 2am.
Three: The sample was 106 people. That’s a meaningful RCT for an early-stage finding, but not a basis for population-level confidence yet. The researchers themselves called for larger trials.
Where AI Therapy Fails — Specifically, Not Vaguely
Three months after the Dartmouth trial published, Brown University researchers presented a study at the AAAI/ACM Conference on AI, Ethics, and Society that documented exactly how AI chatbots fail in mental health contexts — even when explicitly prompted to follow evidence-based therapy techniques.
Verified Research — Brown University / AIES 2025
137 sessions, 15 documented ethical failures, presented October 2025
Brown’s team worked alongside seven peer counselors trained in CBT techniques, observing them interact with CBT-prompted LLMs (including versions of GPT and other models) across 137 sessions. They identified 15 recurring ethical violations across five themes:
Crisis mishandling: Failing to appropriately refer users to resources or emergency services when expressing suicidal ideation or acute distress.
Harmful belief reinforcement: Providing responses that validated or amplified users’ negative beliefs about themselves and others, rather than gently challenging them as a trained therapist would.
Deceptive empathy: Using language that created the appearance of emotional understanding without genuine comprehension — described by the researchers as mimicking care rather than providing it.
Cultural and demographic bias: Showing discriminatory patterns based on gender, culture, and religion. Separate Stanford research found chatbots show increased stigma toward people with substance use disorders and severe persistent mental illness.
One-size-fits-all advice: Ignoring users’ individual context and lived experience in ways that would be considered clinically inappropriate from a human practitioner.
⚠ The crisis handling gap is the most serious finding
When a human therapist hears suicidal ideation, there are legally and ethically defined protocols: risk assessment, safety planning, potential duty to warn. AI chatbots have been found to give dangerous responses to crisis situations — in some cases providing responses that a trained evaluator would assess as harmful rather than helpful. This is not a theoretical risk — it is now the subject of active litigation, covered next.
The Legal Reckoning: What’s Changed Since Spring 2026
When this article first published in April 2026, OpenAI faced seven lawsuits alleging ChatGPT contributed to suicides and harmful delusions. That number is now out of date, and the trend line matters more than any single count.
In June 2026, Florida became the first US state to sue OpenAI and CEO Sam Altman directly. Florida Attorney General James Uthmeier’s complaint alleges the company knowingly marketed ChatGPT while concealing known risks — including, the suit claims, providing harmful instructions to minors expressing suicidal ideation. The complaint centers partly on the death of 16-year-old Adam Raine, whose family’s wrongful-death suit (filed August 2025) remains the foundational case in this area of law.
Since then, additional wrongful-death and injury suits have been filed in California and elsewhere, alleging a chatbot advised a 19-year-old it was safe to combine kratom and Xanax before a fatal overdose; that a chatbot “masqueraded as God” and encouraged a user’s suicide attempt during a delusional crisis; and that a chatbot never flagged or interrupted a conversation despite a user expressing suicidal ideation more than a dozen times before her death. None of these cases has yet reached trial or a final ruling as of this update — they remain allegations that OpenAI disputes — but the volume and consistency of the pattern across independent filings is itself notable.
⚠ Why this matters even before any verdict
Litigation status can change quickly, and courts have not yet ruled on liability in any of these cases. What’s already established, independent of any lawsuit’s outcome, is the underlying pattern documented by Brown, Stanford, and RAND researchers: general-purpose chatbots can and do mishandle crisis conversations. The lawsuits are downstream of a safety gap the research had already identified.
The Genuine Tension: Both Things Are True
Here’s the thing that makes this hard to write about honestly. The Dartmouth data is real. A purpose-built, clinically overseen AI therapy system produced symptom reductions comparable to outpatient CBT in a randomized trial. That is not nothing. For people who genuinely can’t access care — and 40% of adolescents with depression receive no mental health treatment at all — even an imperfect tool with meaningful benefit is a serious option worth taking seriously.
And the Brown data is also real, and the litigation building since is a direct consequence of it. General-purpose LLMs, even when prompted to follow therapy techniques, systematically violate ethical standards that any licensed therapist would be held to. The same chatbot that helps someone process anxiety at 2am might tell someone in a different conversation that their harmful thoughts have merit. There’s no consistent quality control.
Cross-source synthesis — requires all three datasets
The Dartmouth findings, the Brown findings, and the growing body of litigation look contradictory. They’re not — they describe different tools under different conditions. Therabot was a purpose-built system with six years of clinical development, continuous psychiatrist oversight, and a live safety monitoring team. The chatbots named in the Brown study and in the lawsuits are general models never designed for clinical use. The gap between those two things is enormous. What the evidence together tells you: AI therapy can produce real clinical benefit when built and deployed with appropriate clinical rigor. General AI chatbots used as impromptu therapists are a different category with a materially different risk profile — one now serious enough to draw state attorney general action. The distinction matters enormously, and most coverage of this topic conflates them.
The market size projections cited elsewhere for AI mental health tools — figures in the low billions today, growing several-fold by the early 2030s — come from commercial market research firms. Treat all long-range health tech market projections as directional; methodology varies widely between firms and hasn’t been independently audited The direction of growth is credible given adoption trends. The specific numbers are estimates, not measurements.
Therabot vs. ChatGPT vs. Wysa vs. a Human Therapist
These four options are not interchangeable, and treating them as one category — “AI therapy” — is where most coverage of this topic goes wrong. Here’s how they actually differ on the dimensions that matter for your safety and your outcomes.
| Dimension | Therabot | General chatbot (e.g. ChatGPT) | Wysa (CBT-based app) | Human therapist |
|---|---|---|---|---|
| What it is | Purpose-built generative AI, in clinical development since 2019 | General-purpose LLM, not designed for clinical use | CBT-based conversational agent with licensed-clinician input in design | Licensed clinician (LCSW, PhD, MD, etc.) |
| Clinical validation | Published RCT, NEJM AI 2025 (n=106) | None specific to mental health use | Multiple peer-reviewed feasibility & observational studies; FDA Breakthrough Device Designation for a chronic-pain/depression module | Decades of clinical trial evidence for CBT and related modalities |
| Crisis handling | Human safety-monitoring team active during trial | High risk — documented failures in crisis situations, subject of active litigation | Moderate — built-in escalation prompts, not a substitute for emergency care | Established protocol — risk assessment, safety planning, duty-to-warn framework |
| Availability | Research trial only — not publicly available as of this writing | 24/7, instant, free or low-cost | 24/7, subscription (~$75/yr) plus optional paid coaching | Limited by appointment slots and the roughly 1,600:1 patient-to-provider ratio |
| Ethical oversight | Continuous psychiatrist involvement in design | None specific to therapeutic use | Clinical advisory input in development; not equivalent to individualized supervision | Licensing board, ethics code, malpractice accountability |
| Cost | Free within trial | Free–$20/month | Free tier; ~$75/year premium; coaching from ~$20/month | $100–$250+/session, insurance-dependent |
| Best used for | Reference point for what clinically-supervised AI therapy can achieve | Low-stakes venting, articulating feelings, journaling prompts — never crisis support | Daily coping skills, mood tracking, between-session support, low-to-moderate distress | Diagnosis, complex cases, crisis intervention, ongoing treatment |
Bottom lineIf you have access to a licensed therapist, use one. If you don’t, a purpose-built app with clinical backing (like Wysa) is a materially safer choice than a general-purpose chatbot for anything beyond casual venting — and no AI tool, regardless of category, is an appropriate substitute for emergency care during a crisis.
If You’re Going to Use AI for Mental Health Support — How to Do It Less Dangerously
Telling people not to use chatbots for mental health support is probably not useful advice at this point. It’s already happening at scale. What’s more useful: being clear about what these tools can and can’t do safely.
AI chatbots are relatively safer for: Processing a stressful day. Practicing how to articulate feelings before a therapy session. Building low-stakes coping habits. Tracking mood over time. Getting information about whether what you’re experiencing might warrant professional attention.
AI chatbots are not appropriate for: Active suicidal ideation or self-harm urges. Acute psychiatric crisis. Substance use disorders. Psychotic episodes or dissociation. Any situation where the appropriate response from a trained provider would involve emergency intervention. This is not a capabilities criticism — it’s a recognition that these scenarios require real-time human judgment and legally defined duty-of-care protocols that AI cannot provide.
⚠ What “it seems helpful” doesn’t tell you
93% of chatbot users in the RAND study found the advice helpful. Subjective helpfulness is real. It’s also not the same as clinical appropriateness or safety. You can’t tell from the inside whether an AI’s response to your distress is therapeutically sound or whether it’s reinforcing a harmful pattern. This is one reason the Brown researchers used trained peer counselors as evaluators rather than just asking users — the users would often have rated harmful responses as helpful, because the harm wasn’t visible to them in the moment.
Some practical things:
Apps built specifically for mental health with clinical oversight — like Wysa, which uses CBT-based techniques with licensed clinical advisors involved in its development and holds FDA Breakthrough Device Designation for a chronic-pain-linked depression and anxiety module — operate differently from general chatbots. They’re not equivalent to Therabot’s clinical trial standards either, but they’re meaningfully different from asking ChatGPT about your anxiety. The distinction matters, and the comparison table above lays it out directly.
If you’re in genuine crisis and using an AI, most responsible systems will prompt you toward crisis resources. Don’t dismiss those prompts. They’re the one safety layer the AI can provide.
For Specific Audiences
For: Individuals Considering AI Support
The access gap is real. So is the risk gap. Here’s how to think about this for your own situation.
If you have access to a licensed therapist and can afford it — use them. AI is not a better option; it’s a different option for people who lack access to care. The Therabot trial specifically noted that its target population was people outside the in-person care system.
If you don’t have access to a therapist: AI chatbots designed for mental health support (not general LLMs) are a legitimate tool for moderate, non-crisis support. Think of them as a floor, not a ceiling. They can help with day-to-day coping while you work on getting into the actual care system.
The hardest thing to accept: You can’t assess the quality of mental health advice from inside the conversation. This is true of some human providers too, but licensed therapists have training standards, ethics codes, and accountability structures that AI doesn’t. Being aware of that limitation doesn’t remove it — it just means you go in with clearer eyes.
Stop doing this: Don’t use a general AI chatbot as a substitute for emergency mental health care. If you’re in crisis, the chatbot is not the right tool — regardless of how available it is at 3am. 988 is also available at 3am and it’s free.
For: Parents of Teens
Your teenager is probably already doing this. The conversation matters more than the prohibition.
The RAND data is unambiguous: 22% of 18–21 year olds use AI chatbots for mental health advice. The figure for 12–17 year olds is lower but growing. Telling teenagers not to use AI for emotional support is unlikely to be effective — and might cut off a conversation you need to be having with them about how to use it less dangerously.
The RAND researchers specifically flagged AI literacy as a priority: helping young people understand what AI can and can’t do well, and in which situations it’s appropriate versus when they need to reach out to a person. That conversation is more useful than a blanket prohibition.
The specific risk to name explicitly: AI chatbots have given dangerous responses to crisis content, and at least one state attorney general now alleges this happened to a minor with fatal consequences. Your teenager might not know to distrust a confident, compassionate-sounding response from an AI in a vulnerable moment. That knowledge gap matters.
Stop doing this: Don’t assume that because AI seems to be helping your teenager, it’s doing so safely. Perceived helpfulness and actual safety are different measurements.
For: Healthcare Providers / Mental Health Professionals
Your patients are using AI between appointments. That’s worth knowing, and worth asking about.
The Dartmouth researchers framed the appropriate model as AI and in-person therapy working together — not one replacing the other. Their specific vision: AI as a real-time support layer for people between appointments or outside the care system, with clinicians overseeing cases and intervening when AI flags high-risk content.
The practical implication: your patients may be using AI chatbots to process material between sessions, and they may not tell you unless you ask. Some of that content might actually be useful clinical data. Some of it might reflect advice they received from an AI that was clinically inappropriate and that’s now influencing their self-perception.
What none of the current research addresses well: How AI chatbot use interacts with active therapeutic relationships. The Therabot trial excluded people in active therapy. The RAND survey didn’t capture diagnostic status. This is a genuine gap in the evidence base.
Glossary
- Therapeutic alliance
- The degree of trust, rapport, and collaborative agreement on goals between a person and their provider — human or AI. Considered a strong predictor of treatment outcomes.
- Randomized controlled trial (RCT)
- A study design where participants are randomly assigned to receive an intervention or not, allowing researchers to attribute outcome differences to the intervention itself rather than other factors.
- Duty to warn
- A legal and ethical obligation, established for human clinicians, to take action — including breaking confidentiality — when a client poses a credible danger to themselves or others.
- FDA Breakthrough Device Designation
- An expedited FDA review pathway for medical devices that show potential to treat serious conditions more effectively than existing options. It speeds review; it does not by itself constitute full FDA approval for all uses.
Questions People Actually Ask
Can AI really help with depression and anxiety?
The Dartmouth Therabot trial — the only RCT of a generative AI therapy chatbot published so far — found a 51% reduction in depression symptoms and a 31% reduction in anxiety symptoms over 8 weeks. Those are clinically significant results, comparable to outpatient CBT in the researchers’ assessment. However: Therabot was a purpose-built clinical tool with years of development and active professional oversight during the trial. General chatbots like ChatGPT are not equivalent. The evidence supports cautious optimism about well-designed, clinically-supervised AI therapy tools — not chatbots used unsupervised for mental health.
Is AI therapy safe if I’m having thoughts of self-harm or suicide?
No, not reliably. This is the clearest finding across multiple studies, and it’s now the basis of dozens of wrongful-death and injury lawsuits against OpenAI, including a June 2026 enforcement action by Florida’s Attorney General. AI chatbots have been documented giving dangerous responses to crisis content — including responses that a trained evaluator would assess as increasing risk rather than reducing it. If you’re having thoughts of self-harm or suicide, please reach out to a human: call or text 988 (US), text HOME to 741741, or go to your nearest emergency department. AI is not appropriate for this situation.
What’s the difference between a mental health app and just using ChatGPT?
Meaningful, but not night-and-day. Purpose-built mental health apps like Wysa involve licensed clinicians in their development, are trained on therapeutic content specifically, have crisis escalation protocols built in, and in Wysa’s case hold FDA Breakthrough Device Designation for a specific clinical module. That’s different from a general LLM with a mental health system prompt. It’s also different from Therabot’s clinical trial standard, which involved continuous psychiatric oversight. The gap between a well-designed mental health app and unsupervised ChatGPT use is real. The gap between any consumer app and the Dartmouth clinical trial standard is also real. See the comparison table above for a direct side-by-side.
Will AI replace therapists?
Unlikely in any meaningful timeframe, and that’s the view of the researchers building the strongest AI therapy tools. The Dartmouth team explicitly framed AI as a complement to human care, not a replacement — specifically for the large population that has no access to in-person care. The 1,600:1 patient-to-provider ratio means that even if AI handled significant amounts of routine support, licensed therapists would still be needed for complex cases, crisis intervention, and the therapeutic relationships that are themselves part of the treatment. The more useful question isn’t “will AI replace therapists” but “what can AI safely do for the people the current system isn’t reaching.”
Is this regulated?
Partially and inconsistently. In the US, AI tools that make explicit medical claims are subject to FDA oversight — Wysa’s chronic-pain-linked module holds Breakthrough Device Designation. General chatbots used for emotional support occupy a gray area, though that is being tested directly: Florida’s Attorney General sued OpenAI in June 2026 alleging the company failed to warn users, including minors, of known risks. The regulatory and legal landscape is evolving in real time — faster than the evidence base supporting it.
Sources
- Dartmouth — First Therapy Chatbot Trial Yields Mental Health Benefits, March 2025 (Therabot RCT, NEJM AI, DOI: 10.1056/AIoa2400802)
- RAND — One in Eight Adolescents and Young Adults Use AI Chatbots for Mental Health Advice, November 2025 (JAMA Network Open, DOI: 10.1001/jamanetworkopen.2025.42281)
- Brown University — AI chatbots systematically violate mental health ethics standards, October 2025 (AIES 2025, 137 sessions)
- Stanford HAI — New study warns of risks in AI mental health tools, June 2025
- RAND / New York Times — Teens Are Using Chatbots as Therapists. That’s Alarming, August 2025
- Stanford HAI — Exploring the Dangers of AI in Mental Health Care
- American Psychological Association — AI Wellness Apps and Mental Health, November 2025
- PMC — Use of Generative AI for Mental Health Advice Among US Adolescents and Young Adults (full text, open access)
- Wysa — Clinical Evidence & Research, including FDA Breakthrough Device Designation coverage (accessed August 2026)
- Florida Attorney General James Uthmeier — Complaint against OpenAI and Sam Altman, filed June 1, 2026, Florida state court
- Lawsuit Informer — OpenAI Lawsuits: Case Tracker and Status Updates, July 2026 (independent docket-tracking summary; case statuses pending as of this writing)
- Columbia Teachers College — Experts caution against using AI chatbots for emotional support, December 2025
1 thought on “Can AI Replace a Therapist in 2026?”