Regulators Just Told Us Where AI Belongs in Mental Health. It's Not the Therapy Chair.

Buried in a new federal report on AI in mental health services is a sentence that should reframe the entire debate: the strongest near-term opportunities for AI in mental health may be the least visible ones.

That's not a hedge. It's a finding. The report, published by the Substance Abuse and Mental Health Services Administration (SAMHSA) and co-authored by digital psychiatry researcher John Torous, reviewed the evidence across administrative, clinician-facing, and patient-facing applications of AI. The pattern it found is striking: AI's evidence base is strongest in the unglamorous places — documentation, workflow, system-level support — and gets thinner the closer AI moves to direct therapy and crisis response. The tools nobody writes headlines about are the ones that work. The tools everybody writes headlines about are the ones we can't yet trust.

I've spent the past year building in this space, and I read most weeks of mental-health-tech news as noise around that single signal. This week, unusually, everything pointed the same direction.

Start with the regulators. The Wall Street Journal's What's News podcast ran a segment on August 23 titled "The Wild West of AI Therapy Laws," examining the wave of state legislation restricting AI therapy tools. Dartmouth's Nicholas Jacobson — who builds Therabot, one of the few generative AI mental health tools with real clinical trial evidence behind it — noted that these laws "don't often have exemptions for these clinical products that are effective and safe." You can read that as a complaint about blunt regulation. I read it differently: legislatures across the country have looked at chatbots playing therapist and decided, with rare bipartisan speed, that the default answer is no. The burden of proof now sits where it always should have — on the technology, not on the patient.

The same week, a review in DIGITAL HEALTH, drawing on a symposium convened by the UK's medicines regulator, the MHRA, laid out what the next generation of oversight should look like. Its first principle: human oversight remains critical, particularly where AI informs clinical decisions or interacts directly with people seeking mental health support. Its subtler point is one builders should sit with: safety has to extend across the entire AI lifecycle — real-world monitoring, model drift, adverse events — not just a benchmark score at launch. A chatbot that aced its evaluation in January is not the same system in August.

And SAMHSA's report adds the sharpest caveat of all to the "human in the loop" mantra that every AI company, mine included, likes to recite: a human in the loop only counts if the human role is meaningful. A clinician rubber-stamping AI output at volume, without training, time, or accountability, is oversight theater. If we mean clinician-first, the clinician needs real authority over the tool — not a cameo in its marketing.

Meanwhile, researchers at Deakin University's Lifespan Institute published a position paper in The Lancet Psychiatry making a point that the regulatory conversation keeps missing: general-purpose AI systems — ChatGPT, Claude, Gemini — are often a distressed person's first point of contact, ahead of any purpose-built mental health app, and there are still no agreed frameworks governing any of it. Their proposed roadmap focuses on something I find genuinely encouraging: clarifying the optimal sequencing and integration of AI tools into clinical workflows. Not whether AI belongs in mental health care, but where in the care pathway it does the most good and the least harm.

Put the week together and the connective thread is hard to miss. The question has quietly changed. Two years ago the debate was "can AI do therapy?" This week, a federal agency, a national regulator, state legislatures, and a Lancet Psychiatry paper all answered a different question: where does AI belong in a system of care that remains human at its core? And they converged on the same answer. Not in the therapy chair. In the infrastructure around it.

That answer matches what clinicians have been saying all along — the American Psychological Association reported earlier this year that more than a third of psychologists already have patients using AI as a kind of auxiliary mental health professional, and most are far from convinced that's progress.

Here's what I think that infrastructure actually looks like, and why I find the "least visible" framing so clarifying. A therapist sees a patient for one hour out of 168 in a week. The other 167 hours are where life happens — the sleep that collapses, the medication skipped, the conflict that will dominate the next session. Today, almost none of that context reaches the clinician until the patient reconstructs it from memory, weeks later, in fragments. The highest-value AI in mental health isn't a synthetic empathizer. It's the layer that gives a qualified clinician visibility into those 167 hours — and that makes sure the patient is sitting across from the right clinician in the first place, one who knows their context even at a first meeting. Continuity is not a feature of good therapy. It is close to the whole mechanism: the alliance between one patient and one clinician, carried across time.

The companies that will matter in this field five years from now won't be the ones that built the most convincing artificial therapist. They'll be the ones that made real therapists impossible to replace — better informed, better matched to their patients, and present, in some meaningful sense, between the hours they're in the room.

This week, the regulators, the researchers, and the evidence all said the same thing. Builders should listen.