Engagement Is the Business Model. Alliance Is the Treatment.

Dartmouth has spent seven years building a therapy chatbot, and you still cannot download it.

Therabot, developed at Dartmouth's Center for Technology and Behavioral Health, is among the first generative AI mental health chatbots to show results in a published trial. As the Valley News reported this week, it exists only inside approved research studies, and psychiatrist Michael Heinz, its lead clinical investigator, cannot say when, or whether, it will reach the market. Meanwhile, apps calling themselves AI therapists are downloadable today, some borrowing the Therabot name with no connection to Dartmouth.

That gap, between the careful clinician-led version of this technology and the engagement-optimized version already in millions of pockets, was the story of this week. It played out in four places: a federal courtroom in Oakland, statehouses in Sacramento and Montpelier, a congressional letter to the VA, and a research lab in New Hampshire. Each is wrestling with the same question. Who holds the clinical relationship, and who is accountable when it breaks?

Start in Oakland. On Tuesday, a federal trial opened after the Ninth Circuit refused to delay it, consolidating claims from 29 state attorneys general that Meta and TikTok designed their products in ways that harmed young users (Medical Daily, Aug 12). The shape of the allegation is what stands out. It is about design, not content: the claim that features were engineered to maximize engagement among minors, and that the companies understood the consequences. Nothing has been proven, and the science on social media and teen mental health remains contested. But the legal system is now stress-testing a business model, and the model on trial is engagement for its own sake.

That same model is what California lawmakers are trying to wall off from care. Senate Bill 903, authored by Sen. Steve Padilla and covered by CalMatters earlier this month, would ban companies from advertising chatbots as therapy and prohibit AI from making therapeutic decisions without review by a licensed professional. The line Padilla draws is precise: AI can handle administrative work and support clinicians, and it cannot practice psychotherapy. Le Ondra Clark Harvey of the California Behavioral Health Association put the stakes plainly: "The difference between a licensed clinician and an automated response is not technical. It can be life altering." Vermont has drawn a similar line. Its H.816, signed in June, requires a licensed professional to review AI-generated mental health advice before it reaches a patient.

I find it telling that the strongest case for AI in mental health this week came from someone building it under exactly those constraints. Heinz's team trained Therabot from scratch on data matched to therapeutic best practices, hosts it on university servers, and holds it to a design principle many consumer apps invert. "We are never optimizing toward engagement for the sake of engagement," he told the Valley News. "It really is optimized toward measures of human wellbeing." His use case is worth sitting with. A patient leaves a session with an assignment and, instead of facing the week alone, has a clinically supervised tool alongside them for the exposures and exercises. That is the space between sessions, made visible.

The fourth story explains the urgency: the humans are stretched thin. Sixty-two House Democrats gave the VA until August 14 to produce facility-level data on mental health staffing, citing the department's own figures. More than 300 psychologists and roughly 700 social workers have been lost since the start of the current administration, and an inspector general found that 57% of facilities report severe psychology staffing shortages (Medical Daily, Aug 10). Buried in that letter is the detail I keep returning to. Trauma-focused therapies run as structured multi-week protocols that depend on continuity. A veteran reassigned mid-protocol often restarts from zero, recounting their trauma to a stranger. Headcount numbers miss this. Continuity is part of the treatment.

Put the four together. Courts are interrogating engagement-driven design. Legislators are ring-fencing the practice of therapy for licensed clinicians. The clinical workforce is too depleted to pretend the status quo is fine. And the most credible AI research accepts clinical accountability as a design constraint rather than a compliance burden.

The wrong conclusion is that technology has no place here. The access gap is real. Heinz is right that a chatbot does not keep office hours or a waitlist, and the VA data shows what happens when the supply of clinicians erodes. The right conclusion is narrower. The value of AI in mental health scales with its proximity to a real clinical relationship, not its distance from one. Tools that extend a therapist's sight into the week between appointments, and that protect the therapeutic alliance instead of substituting for it, pass the test every arena applied this week. Tools that quietly become the relationship do not.

The engagement economy has had its decade with our attention. Care deserves a different architecture, one where a qualified human is accountable for the person on the other end and the software works for that human. That is the direction this week pointed. I do not think it is turning back.