AI Therapy: What It Is, How It Works & Its Limits

What is AI therapy?

AI therapy usually means a text or voice conversation with software designed to provide emotional support. The AI responds to what you share with questions, reflections, exercises, and suggestions that may draw on approaches such as CBT or DBT. Some products are built specifically for mental health. Others are general chatbots being used that way by their users.

The name is convenient, but it can also be misleading. Therapy is a regulated clinical service delivered by a licensed human. AI can offer support, structure, and practice; it cannot accept clinical responsibility for your care. That gap is especially important when comparing a purpose-built mental health app with a general chatbot. We explain the latter in our piece on using ChatGPT for mental health.

How does AI therapy work?

You type or speak about what is happening, and the system generates a response from the context of that conversation. A purpose-built app may add structured exercises, mood tracking, memory between sessions, and rules for handling crisis language. The useful part is not that the AI has feelings. It is that a structured conversation can help you name a thought, see a pattern, rehearse a difficult conversation, or practise a coping skill in the moment you need it.

Quality varies widely. A warm tone is easy to generate; safe boundaries, privacy, mental-health-specific design, and appropriate escalation are harder. Those are the features worth comparing—not whether a chatbot sounds human.

What makes therapy, therapy

Therapy is not defined by talking about feelings. It is defined by who is responsible for what happens next. A licensed therapist can diagnose you, build a treatment plan, adjust it when it stops working, coordinate with a doctor, and be held professionally accountable if they harm you. They carry a legal duty of care. You can file a complaint with a licensing board and end their career.

No AI has any of that. An AI cannot diagnose you, cannot be sued for malpractice, and cannot sit in the silence after you say the thing you have never said out loud and read what your face is doing. Whatever the marketing says, the accountability structure that makes therapy a profession simply does not exist for software.

Regulators have started drawing the same line. In August 2025, Illinois passed the Wellness and Oversight for Psychological Resources Act, which prohibits offering therapy or psychotherapy services through AI unless a licensed professional performs them, with fines up to $10,000 per violation. Other states are moving in the same direction. The era of apps casually calling themselves "your AI therapist" is ending, and honestly, good.

Does AI therapy actually work?

Here is the half the skeptics leave out. The first randomized controlled trial of a generative AI mental health chatbot, Dartmouth's Therabot study, was published in NEJM AI in March 2025. Participants with major depressive disorder saw an average 51% reduction in symptoms over four weeks. Those with generalized anxiety saw a 31% reduction. Participants also reported an ability to trust and communicate with the system at a level researchers compared to working with a human professional.

Before anyone frames that as "AI beats therapists," some honesty about the fine print. It was one trial with 106 people using the tool. It ran four weeks. Clinicians monitored the interactions. And Therabot was purpose-built for mental health over years, which is exactly the point: the encouraging evidence comes from tools designed for this work, not from general chatbots doing an impression of empathy. Wysa, the veteran app in this space, has peer-reviewed studies on its own product for the same reason.

So the fair summary is: purpose-built AI support can measurably help with depression and anxiety symptoms, the evidence base is young, and none of it makes the tool a therapist.

A broader 2026 systematic review of 16 randomized controlled trials was more cautious: only a minority of the included studies found significant improvements in anxiety or depression symptoms. That does not mean AI support is useless. It means one successful product or trial cannot prove that every tool marketed as AI therapy works.

Where AI genuinely helps

Most of mental health does not happen in a therapist's office. It happens on a Tuesday night when the overthinking starts, in the parking lot before a hard conversation, in the weeks between sessions, or in the years before you can afford a session at all. That is the territory where AI support earns its place:

• It is there at the moment it is needed. Waitlists in many places run months. An app answers in seconds, at any hour, which matters because spirals do not schedule themselves.
• It removes the audience. Plenty of people rehearse for weeks before saying something true to a human. Saying it first to something that cannot judge you lowers the cost of saying it at all.
• It can teach real skills. Approaches like CBT and DBT are structured enough that software can walk you through them honestly, and practice between moments of crisis is where those skills actually form.
• It remembers. A good app carries context from Tuesday into Thursday, so you are building on something instead of starting over each time.

This is the job we built sokoon for. Conversations informed by CBT, DBT, and clinician-reviewed dialogues, by text or voice, with memory across sessions, mood tracking, and encryption. It is a support tool with real structure behind it. It is not therapy, and it tells you so.

How much does AI therapy cost?

Prices range from free plans to monthly subscriptions. Free access may mean a short trial, a message limit, or a smaller set of features, so check what happens after the first conversation and whether a credit card is required. sokoon has a permanent free plan with one text session each day and no card. Our free AI therapy guide explains exactly what is included and where the paid plans begin.

How to choose an AI therapy app

• Purpose: Prefer a tool designed for mental health over a general chatbot with a supportive prompt.
• Privacy: Check whether conversations are encrypted, sold, or used to train models.
• Safety: Look for clear crisis boundaries and a route to human help.
• Method: Ask what structure sits behind the conversation, such as CBT or DBT exercises.
• Limits: Avoid products that promise diagnosis, treatment, or a replacement for licensed care.

If you are comparing products described as an AI therapist, read our AI therapist guide. If you want to compare actual apps, see our AI mental health app comparison.

When you should skip AI entirely

If you are in crisis, thinking about harming yourself, or experiencing things others say are not there, an app is the wrong tool. Not because you do not deserve support, but because you deserve more support than software can carry. Call or text a crisis line, tell a person you trust, or go to an emergency room. An honest AI app should tell you exactly this and point you to real resources the moment the conversation gets there. If an app instead keeps you chatting, close it. We wrote more about that failure mode in can AI conversations harm your mental health.

The same goes for diagnosis. If you suspect depression, an anxiety disorder, BPD, or anything else with a name, an AI can help you organize what you are noticing, but only a clinician can tell you what it is and what treatment fits.

Why we refuse to call sokoon therapy

Partly because it is illegal in a growing number of places. Mostly because it is untrue, and untrue is a bad foundation for a mental health product. When an app calls itself your therapist, it writes a check it cannot cash: it borrows the trust that a century of clinical practice built, without the training, oversight, or accountability that earned it. Then, at the worst possible moment, the user discovers the difference.

Calling it what it is costs us a shinier headline. It buys something better: you know exactly what you are getting, and what you are getting is still worth a lot. A private place to say the unsayable, structure drawn from approaches that actually work, and a tool that knows where its job ends and a professional's begins.

Quick answers

Does AI therapy work? Some purpose-built tools have encouraging trial results, but the broader evidence is mixed and does not apply to every app.

Is it real therapy? No. No diagnosis, no treatment plan, no licensed human, no legal accountability.

Is it safe? Reasonably, if the app was built for mental health, has crisis protocols, and protects your data. Never as a crisis service.

Can it replace a therapist? No. It works best between sessions, or as a first step when therapy is out of reach. Our guide to AI for mental health covers how to evaluate any app you are considering, including ours, and if you came here searching for an AI therapist, we wrote the honest version of that page too.