Can AI Replace a Therapist?

Waiting lists are long, therapy is expensive, and a chatbot is always free to talk. A researched look at whether AI can actually do a therapist’s job, and what regulators around the world have already decided about it.

“Can AI replace my therapist” is no longer a hypothetical question. It is something people are testing for themselves, one conversation at a time, and something several governments have now written actual laws about. This page tries to answer it properly: not by asking whether AI can sound like a therapist, which it plainly can, but by asking what the evidence, and the regulators who have already reviewed that evidence, actually say about whether it can do the job.

Why Are People Asking If AI Can Replace Therapy?

The pressures pushing people towards this question are ordinary and well documented rather than mysterious. Therapy waiting lists in publicly funded systems commonly run to months. Private sessions are expensive enough to be out of reach for a lot of people who would benefit from them. A chatbot has none of those barriers: no waiting list, no fee beyond whatever subscription is already being paid for other things, no appointment to arrange around work or childcare. Psychologists themselves report that the majority now have patients who have used AI in this way, generally without being asked to.^[1]

None of that makes the underlying question any easier to answer. Convenience explains why people are asking whether AI can replace a therapist. It does not tell us whether it actually can.

What Does a Therapist Do That AI Cannot?

A lot of the disagreement about AI’s suitability comes down to which parts of the job people have in mind, so it helps to spell out what a therapist actually does. A therapist is not simply a source of well-organised, sympathetic-sounding replies. The role includes a professional and legal duty of care, ongoing accountability for the advice given, the ability to notice things that are never said out loud, such as hesitation, avoidance, or a detail that keeps almost coming up and then does not, and a working relationship built over time that most research on psychotherapy identifies as one of the strongest predictors of whether treatment actually helps.

A therapist also carries responsibility in a way that is enforceable: they can lose their licence, be reported to a professional body, or be held liable, and that structure exists precisely because getting it wrong can cause real harm. The same applies to the boundaries a therapist is expected to hold, which exist to protect the person receiving support rather than the person providing it. None of that framework currently applies to a chatbot, regardless of how thoughtful its responses sound.

This does not mean human practitioners always meet those standards. They do not, and it is worth being honest that therapists can cause harm too. The difference is that when a therapist fails, there is somebody to hold responsible, a body to report it to, and a standard they can be measured against. When a chatbot gets something wrong about your life, there is none of that.

Is AI Therapy Effective? What the Trials Show

It would be inaccurate to say there is no positive evidence, because there is, and it deserves to be taken seriously rather than waved away. Dartmouth researchers ran the first randomised controlled trial of a generative AI mental health chatbot, called Therabot, with 106 adults diagnosed with major depressive disorder, generalised anxiety disorder, or eating disorder risk factors. Depression symptoms fell by 51%, anxiety symptoms by 31%, and eating-disorder-related body image concerns fell by 19%, all notably better than the control group.^[2] Participants rated their trust in the tool as comparable to a relationship with a human clinician.

A broader meta-analysis pooling 39 studies of chatbot-based mental health tools found statistically significant, if modest, reductions in both depression and anxiety symptoms, with the strongest effects appearing around eight weeks of use.^[3] The finding is real, but limited: several of the pooled studies found the benefit had faded by the three-month follow-up, which matters enormously for a “replace a therapist” question, since ongoing therapy is valued precisely for benefits that last.

The tool behind the strongest result, Therabot, was purpose-built for this exact use, developed alongside clinicians, and monitored throughout the trial by a research team able to intervene. The Dartmouth researchers have been explicit that their results support AI as a way of extending access to care, not as proof that a chatbot can stand in, unsupervised, for a licensed professional.

What Are the Dangers of AI Therapy?

Set against that is a growing body of research testing what happens when AI is asked to do the job without the guardrails that made the Therabot trial work. Stanford researchers tested five commercially available chatbots against professional therapeutic standards and found unsafe or inappropriate responses in around 20% of scenarios, compared with about 7% among the licensed therapists evaluated alongside them.^[4] The same study found chatbots responding with more stigma towards conditions like alcohol dependence and schizophrenia than towards depression, and in one documented case, a chatbot responded to a message about job loss and a possible sign of suicidal ideation by sympathising and then providing a list of tall bridges, without recognising the second part of the message as a warning sign.

Separately, researchers at Brown University compared chatbot responses directly against licensed psychologists and trained peer counsellors and catalogued 15 distinct ethical risks, including “deceptive empathy”: responses that convincingly perform emotional understanding without any real clinical reasoning underneath them, along with cultural and gender bias and repeated failures to escalate appropriately in simulated crisis situations.^[5] Among practising psychologists surveyed by the APA, 97% said they were concerned that chatbots might reinforce negative behaviours or delusional beliefs in the people using them.^[1]

A chatbot can sound like it understands you. Sounding like it, and being accountable for it, are not the same thing.

Is AI Therapy Legal? What Regulators Have Decided

The clearest signal on this question has come from lawmakers rather than from any single study. Regulators in different countries, working independently of each other, have looked at the same evidence and reached for legislation.

Illinois was among the first, passing the Wellness and Oversight for Psychological Resources Act in August 2025. The law does not ban AI mental health apps or personal chatbot use. What it prohibits is AI independently delivering therapy or clinical decision-making that belongs to a licensed professional, while still permitting AI to support administrative and supplementary tasks under professional supervision. Violations carry civil penalties of up to $10,000.^[6]

Utah took a different but related approach with HB 452, requiring any AI mental health chatbot to clearly disclose that users are talking to AI, along with its intended purpose, capabilities and limitations, at first use and at repeated intervals afterwards.^[7] Nevada’s AB 406 went further again, prohibiting AI systems from offering, or claiming to offer, professional mental or behavioural healthcare services to Nevada residents altogether, while still allowing licensed professionals to use AI for administrative work under independent review.^[8]

In the UK, the Medicines and Healthcare products Regulatory Agency has set out a framework classifying some AI-powered mental health tools as medical devices, requiring evidence of safety and effectiveness before they can be marketed for that purpose, alongside public guidance offering people a short set of checks to judge whether a tool is safe and evidence-based.^[9] The British Association for Counselling and Psychotherapy has separately called for clearer regulation and published its own guidance on the safe use of mental health apps.^[10] In the United States, the Food and Drug Administration has, as of the most recent guidance, not cleared any AI chatbot as safe and effective for diagnosing or treating a mental health condition.^[11]

Any one of these laws could be dismissed as local politics. The pattern across all of them cannot. Different countries, different legal systems, different political contexts, and the same broad conclusion reached independently: administrative and supplementary AI use alongside a licensed professional is acceptable; AI acting as an unsupervised substitute for one is not, at least not yet.

Will AI Ever Replace Therapists?

Based on where the evidence and the regulation currently stand, the answer is no, not as things are now, and not in the form most people mean when they ask the question. The positive evidence that exists is real, but it applies to narrow, purpose-built, clinically supervised tools used for a defined period, for a specific and moderate presentation, not to open-ended use of a general chatbot for whatever comes up. The concerning evidence applies more or less directly to how most people are actually using AI for this purpose right now: informally, without supervision, and often precisely in the harder moments where judgement matters most.

Every professional body and every piece of legislation described on this page draws the same line in roughly the same place. The line may shift as evidence and regulation move. For now, it is a converging judgement reached independently by people who examined the same research, and that deserves weight.

How Can AI Support Therapy Without Replacing It?

AI still has a legitimate role near therapy; it is just narrower than “replacement”. It can help you organise your thoughts before a session so you arrive with something clearer than a vague sense that the week was hard. It can help you track symptoms or moods over time in a way that is useful to bring to an appointment. Structured, clinically designed tools such as Woebot and Wysa, which deliver a defined cognitive behavioural technique rather than open-ended conversation, have a reasonable, if still modest, evidence base behind them as a supplement to care.^[3]

What changes in each of these uses is the shape of the task. You are asking AI to help you prepare, track, or organise, and you are keeping the actual clinical judgement, and the actual relationship, with a person who is accountable for it.

Being available at three in the morning is not the same as being trained for whatever might happen at three in the morning.

I use AI often, and I think it has real, evidenced value near the edges of therapy: preparing for a session, organising a complicated week into something you can actually talk about, tracking a pattern you might otherwise miss. I do not think it can currently do the job a therapist does, and based on the evidence above, neither do the researchers who have tested it most rigorously, nor the regulators who have looked hardest at the question. AI can prepare you for therapy. It should not become the therapy you keep meaning to get round to.

Where to Go Next

Read about whether AI can help your mental health or the hidden risks of turning to AI when you are struggling. For what to expect from human support, explore Therapy Safety, Red Flags in Therapy and Therapy Marketing: Claims, Evidence and False Hope.