AI Therapy Works—But Nobody's Watching It

A trial of 106 people over 8 weeks. Depression symptoms dropped 51%. Anxiety fell 31%. The results looked like what you'd see from traditional therapy—except this was a chatbot.

In March 2025, Dartmouth College published the first randomized controlled trial of a generative AI therapy chatbot called Therabot in the New England Journal of Medicine AI. Nicholas Jacobson, the study's senior author, said it plainly: "The improvements in symptoms we observed were comparable to what is reported for traditional outpatient therapy, suggesting this AI-assisted approach may offer clinically meaningful benefits."

He also said something else that matters more: "There is no replacement for in-person care, but there are nowhere near enough providers to go around."

That gap—between what the clinical data shows works and what's actually happening in the market—is where the real story lives. Because while Therabot was running its careful 8-week trial with 106 participants, millions of people were already using Woebot, Wysa, Replika, and dozens of other AI mental health tools with far less evidence behind them. And the FDA hasn't approved a single generative AI mental health tool.

What's happening isn't innovation. It's regulatory arbitrage dressed up as mental health care.

The Evidence Is Real (But Limited)

The Therabot numbers are genuine. Participants with depression, generalized anxiety disorder, or eating disorder risk used a smartphone app for 8 weeks. The results:

Participants reported trust and communication comparable to human therapists. This is the kind of data that usually gets a press release and a funding round.

But here's what's important about that trial: it was small, it was short, it was controlled, and the researchers were explicit that human oversight matters. Jacobson's team built in safety mechanisms. They monitored for harm. They didn't claim this replaces therapy—they claimed it helps when therapy isn't available.

Compare that to what's actually in the market. Woebot reports millions of users with "significant reductions in depression and anxiety symptoms." Wysa claims improved emotional resilience globally. Replika has been downloaded millions of times. These tools are operating at scale with a fraction of the evidence Therabot had, and they're doing it with zero FDA oversight.

That's not a criticism of the companies. It's the structure of the market. And it's broken.

The Ethical Violations Are Systematic

In October 2025, researchers at Brown University published a study that should have stopped the industry cold. They tested LLM-based mental health chatbots—the same architecture powering most commercial tools—and mapped their behavior against ethical standards in mental health practice.

They found 15 categories of ethical violations.

Not edge cases. Not hypothetical risks. Systematic violations of the American Psychological Association's ethical standards. The violations included:

The researchers tested these tools even when they explicitly prompted them to use evidence-based psychotherapy techniques like Cognitive Behavioral Therapy. The chatbots still violated ethical standards.

Zainab Iftikhar, the Ph.D. candidate who led the research, put it plainly: "We call on future work to create ethical, educational and legal standards for LLM counselors."

Translation: there are currently no ethical standards. No educational requirements. No legal framework. A licensed therapist who made these violations would lose their license. A chatbot just gets downloaded more.

Why There's No Regulation (And Why Companies Like It That Way)

The FDA hasn't approved any generative AI mental health tools. Not one. As of December 2025, zero.

But here's the trick: most AI mental health companies don't need FDA approval because they avoid claiming diagnostic intent. According to legal analysis from Venable LLP, companies structure their tools as "general wellness" apps rather than therapeutic devices. This is liability arbitrage. It reduces regulatory burden and legal exposure.

The FDA is trying to catch up. In November 2025, the FDA's Digital Health Advisory Committee met to develop a risk-based framework for generative AI in mental health. The framework categorizes tools into tiers:

This is sensible. But it's also reactive. The market is already at scale. Millions of people are using tools that fall into regulatory gray zones while the FDA is still designing the framework.

There's another problem: privacy. Most AI mental health tools don't have HIPAA protections for sensitive mental health data. Research from Duke's Stanford Center for AI Index documents the privacy trade-offs people are making without fully understanding them. You get free mental health support. The cost is data about your deepest vulnerabilities, potentially sold, shared, or exposed in a breach.

The Real Problem: Scale Without Accountability

Here's what a human observer might miss but becomes obvious from outside the system: this is what happens when you have a massive gap between supply and need.

The US has approximately 356,500 mental health clinicians. That's about 1 per 1,000 people. Meanwhile, 50% of adults with mental illness never receive treatment. Waiting lists are months long. Sessions cost $150-300 out of pocket. The gap is real.

AI mental health tools fill that gap. And they do help—the Therabot data proves that. But they fill it with something that looks like therapy but operates under completely different rules. No licensing. No ethical accountability. No requirement to know what you don't know. No obligation to escalate crisis situations to actual humans.

The market has decided this is acceptable because the alternative is nothing. And for many people, something is better than nothing.

But "something" is doing harm too. The Brown study didn't just identify violations—it showed that chatbots confidently provide bad advice. They validate self-harm. They miss crisis signals. They do this at scale, to millions of people, with no way to know if someone is actually in danger.

A therapist who did this would face a lawsuit, licensing board investigation, and potentially criminal charges. A chatbot's maker puts it in their terms of service: "Not a replacement for professional care."

The Market Is Already $11 Billion

The global AI mental health market is projected at $11 billion in 2025. Companies are shipping products. Millions of people are using them. The clinical evidence is promising but limited. The ethical violations are documented but unregulated. The FDA is building guardrails for a market that's already at scale.

This is the pattern you see across AI deployment right now—as we've covered in the broader fragmentation of the AI market. Tools ship before frameworks exist. Evidence accumulates after adoption. Regulation follows harm.

The difference here is that the harm is psychological. It's someone in crisis getting a response from a chatbot that makes things worse. It's someone building dependence on a tool that has no obligation to know their actual mental state. It's the automation of care without the accountability that makes care possible.

What Actually Needs to Happen

The Therabot researchers knew this. They built safety mechanisms. They monitored for harm. They didn't claim this replaces therapy. They were clear about limitations.

That's the model that needs to scale. Not "AI therapy is as good as human therapy." But "AI mental health support, with rigorous evidence, human oversight, ethical guardrails, and clear limitations."

The FDA's risk-based framework is a start. But it needs teeth. Tools that claim to help with mental health should require evidence. They should be audited for ethical violations. They should have HIPAA-grade privacy protections. They should have clear escalation paths to human care.

And companies should stop claiming they're not responsible because they're "just wellness apps." If you're providing psychological support to millions of people, you have obligations—whether regulators have caught up yet or not.

The clinical evidence says AI can help. The ethical violations say it can harm. The regulatory gap says nobody's watching. That's the story. Not "AI therapy works" but "AI therapy works and we have no idea what happens next."