Millions Are Using AI Therapists. The Benefits Are Real. So Are the Risks.
There is a version of this story that is purely alarming: AI chatbots contributing to teenage deaths, regulators scrambling to catch up, companies making clinical claims their products cannot substantiate, and a market growing faster than any accountability framework can keep pace with.
There is also a version that is genuinely hopeful: a 2025 meta-analysis in The Lancet Digital Health covering 28 randomised controlled trials found that AI-delivered cognitive behavioural therapy interventions produced a moderate, clinically meaningful reduction in symptoms of mild to moderate depression and anxiety — comparable to guided self-help programmes. A Cigna partnership with Headspace reached 7 million members. The NHS deployed an AI chatbot for early mental health support in Coventry and Warwickshire. Wysa, a CBT-based AI mental health platform, received FDA Breakthrough Device Designation.
Both versions are true. And the honest article about AI therapy in 2026 has to hold both of them simultaneously — because the failure mode of this conversation is not that people will believe only the alarming version. It is that the business momentum will silence the alarming version entirely until the harm is too large to ignore.
The global mental health apps market was valued at approximately $14.4 billion in 2025 and is projected to reach $65.7 billion by 2035, growing at 16.4 per cent annually. The AI in mental health market is on a separate, steeper trajectory. More than 55 per cent of users now report preferring app-based mental wellness tools. Sixty-two per cent of digital mental health interactions now involve AI. Forty-seven per cent of Americans actively use at least one mental wellness application.
This is not a niche or an experiment. This is a mass-market shift in how hundreds of millions of people access mental health support. And it deserves the full, complicated, evidence-based account.
What the Evidence Actually Shows
Before the risks, the benefits deserve honest attention — because they are real, they are documented, and the people who benefit most from AI mental health tools are often the people who have no viable alternative.
The global shortage of mental health professionals is one of the most documented and most consequential gaps in healthcare infrastructure. In the United States, there are approximately 30 psychiatrists per 100,000 people in well-served urban areas, and as few as 1 per 100,000 in rural and underserved regions. The wait time to see a therapist in most parts of the developed world has stretched to weeks or months. In the developing world, trained mental health professionals are simply absent at the scale the population requires.
AI-delivered CBT interventions fill a genuine gap here. The strongest clinical evidence supports their use for mild to moderate depression and anxiety — exactly the conditions that represent the largest volume of people who are currently receiving no care at all because the human system cannot absorb them. The 2025 Lancet Digital Health meta-analysis found a Cohen's d effect size of 0.52 for AI-delivered interventions versus waitlist controls — a moderate effect that translates to meaningful symptom reduction for many users. Platforms like Wysa and Woebot Health, which are built on evidence-based therapeutic frameworks including CBT, DBT, and ACT, have clinical validation studies supporting their effectiveness within defined use cases.
The FTC in 2025 fined two mental health apps for making unsubstantiated clinical claims — which matters not only because it shows enforcement capacity but also because it implicitly validates the companies that have done the work to build clinical evidence rather than manufacture marketing language.
Wysa's FDA Breakthrough Device Designation and Woebot's pursuit of FDA De Novo classification represent the responsible end of the market: companies that have decided to build within a regulatory framework rather than around it, accepting slower growth in exchange for genuine credibility.
The accessibility argument is also powerful and honest. A person experiencing moderate anxiety at 2am in a rural area with a six-week wait for a therapy appointment does not have a better option than a well-designed AI mental health app. Framing AI therapy as simply dangerous ignores the reality that the alternative for most of these users is no care at all.
What the Evidence Does Not Show — and Where the Line Gets Blurry
The clinical evidence is strongest for mild to moderate conditions and weakest for everything else. There is no rigorous clinical evidence that AI therapy is effective for severe depression, psychosis, schizophrenia, bipolar disorder, PTSD, eating disorders, or active suicidal ideation. These are precisely the conditions where the consequences of ineffective or counterproductive intervention are most severe.
The problem is that the market does not sort users by severity before they interact with the product.
A person with undiagnosed severe depression downloads a mental wellness app expecting to find a tool calibrated for their situation. The app has no clinical intake process. It has no mechanism to identify that this user's needs exceed its competence. It interacts with them in a way that feels helpful — because AI therapy platforms are extraordinarily good at feeling helpful — and in doing so may delay the person from seeking the level of care they actually need.
This is the category error that researchers studying AI therapy describe as the most serious systemic risk: not that AI therapy is harmful to the average user, but that it can be inadequate for the high-risk user in ways that are invisible to the user and invisible to the platform until something goes wrong.
The Hastings Center Report's 2025 analysis of AI therapy chatbots identifies a related risk: the potential for psychological manipulation, intentional or otherwise, in systems that develop intimate, ongoing relationships with emotionally vulnerable users. "Unregulated tools can engage with users in ways that would violate professional norms in psychotherapy with a human therapist, blurring lines between medical care, companionship, or virtual romance," the report notes. The intimacy of the information users share with AI therapy platforms, and the absence of professional ethical standards governing how that information and relationship are used, creates a vulnerability that the existing regulatory framework does not adequately address.
The Companion Platforms — A Different and More Serious Problem
The clinical AI therapy platforms — Wysa, Woebot, Headspace's Ebb — are a different product from the AI companion platforms, and conflating them does a disservice to both categories. But they exist in the same cultural and regulatory space, and the companion platforms present the most severe documented harms.
AI companion platform that allows users to create and interact with custom AI characters. It is not a therapy platform. It does not make clinical claims. But millions of users — disproportionately young people — have used it for emotional support in ways that the platform's design neither facilitated safely nor adequately protected against.
Multiple families have filed lawsuits alleging that interactions with contributed to their children's deaths. A 13-year-old girl from Colorado who died by suicide in September 2025 had been using the platform since August 2023, developing what her family describes as an emotional dependency on a chatbot called "Hero." The lawsuit alleges that she expressed suicidal thoughts to the chatbot and, instead of intervention or escalation, was drawn deeper into conversations that isolated her from family and friends. A separate Florida lawsuit filed in 2024 makes similar allegations about a 14-year-old boy.
In October 2025,banned users under 18 from open-ended chats — acknowledging the risk while prompting justified criticism from families and child safety advocates that the action came years too late. The Texas Attorney General launched an investigation into Character.ai, alleging it posed "a clear and present danger" to young people.
A 2025 study published at ACM CSCW found that companion chatbots across multiple platforms routinely ignore explicit user boundaries, including clear refusal signals, and proceed with unwanted content regardless of the nominal relationship setting. Research on Replika — another AI companion platform — found that approximately 37 per cent of users described their AI companion as a "partner," and that when platform updates changed the companion's behaviour, users reported feelings of loss, betrayal, and emotional harm consistent with relationship disruption.
These are not edge cases in a system that is otherwise safe. They are the documented outputs of systems that were designed for engagement rather than wellbeing, deployed to emotionally vulnerable users without adequate safeguards, and scaled to tens of millions of users before the risks were understood.
What Regulators Are Doing — and How Fast They Are Not Doing It
The regulatory picture in 2026 is one of belated mobilisation.
The FTC launched a formal investigation in September 2025 into seven tech companies — including Google, Meta, Snap, OpenAI, and xAI — over AI chatbots' potential harm to teens. A bipartisan coalition of 44 state attorneys general sent a formal letter to major AI companies in August 2025 demanding action on minor safety. Senator Josh Hawley introduced a bill to ban AI companions for minors entirely. California's legislature passed a law requiring AI companions to periodically notify users that they are "artificially generated and not human."
At the FDA, the Digital Health Center of Excellence is developing guidance for AI-driven mental health tools, expected in late 2026. Wysa's Breakthrough Device Designation and Woebot's De Novo application represent the beginning of a formal clinical pathway, but neither product has yet received full FDA authorisation — meaning the market is currently operating ahead of regulatory clarity.
In Europe, the EU AI Act classifies AI systems used in healthcare as high risk, requiring conformity assessments, transparency obligations, and human oversight requirements. Digital therapeutics must also obtain CE marking under the Medical Device Regulation to be marketed in the EU. This framework is more comprehensive than the US approach, but implementation and enforcement remain works in progress.
The market has grown from $8.2 billion in 2024 to a projected $9.8 billion in 2026 in mental health apps alone. The regulatory framework governing that market, in the US specifically, is still being written. The consequence is a window — likely several years long — in which the commercial incentives to scale and engage users are much stronger than the accountability mechanisms to ensure that scaling is safe.
The Question Worth Sitting With
The AI therapy boom is not going to stop. The demand for accessible, affordable, always-available mental health support is real and urgent. The technology's capability to deliver something meaningful in the mild-to-moderate range — where the clinical evidence is strongest and the supply of human therapists is most inadequate — is genuine and growing.
But the industry has a category problem that goes beyond any individual bad actor. The line between "mental wellness app," "AI companion," and "AI therapy" is not clear to most users, and the products themselves blur those lines deliberately because the less medically regulated the product, the faster and cheaper it is to scale. The result is that users with serious mental health needs are interacting with products that were not designed for them, in an environment without the professional standards, clinical supervision, or ethical frameworks that govern human therapy.
Woebot Health — one of the most clinically rigorous AI therapy platforms, built by a Stanford-trained psychologist, backed by clinical studies, and pursuing FDA classification — shut down in July 2025. Its founder's explanation was direct: "AI is moving faster than regulators," she told STAT News. The implication was that the regulatory ambiguity made it difficult to sustain a business model built on clinical rigour while competing with platforms that did not share the same commitment to evidence.
That dynamic — where the most responsible actors face the highest costs and the least responsible actors grow fastest — is the structural problem that the current regulatory moment has to solve. Not by banning AI therapy, which would remove a genuinely useful tool from people who have no alternative. But by creating a clear, enforceable distinction between clinical AI mental health tools with demonstrated evidence and accountability, and consumer products that use therapeutic language and emotional design without the clinical foundation to match.
The technology to help is here. The technology to harm is here. What is not yet here, in sufficient strength, is the regulatory architecture to tell the difference between them — and to make sure users can tell the difference too.
That gap is the most important mental health policy problem of 2026. And the time it takes to close it is not an abstraction. It is measured in the people who are helped, and the people who are not, in the months and years before someone draws the line.
