The Chatbot-Therapist Era Is Ending. What Replaces It Matters More.

Last November, a Harvard psychiatrist told Congress that OpenAI itself had reported over one million users per week having ChatGPT conversations containing explicit indicators of potential suicide planning. One million. Per week. On one platform.

I keep returning to that number, because this week made clear that three groups who rarely agree on anything — psychiatrists, state legislators, and venture capitalists — have quietly converged on the same conclusion about what to do about it. The chatbot-as-therapist experiment is being dismantled in real time. And the question that matters now isn't whether AI has a role in mental health care. It's which role.

Start with the clinicians. A Medical Daily piece this week pulled together the first half of 2026's research record, and it reads like a verdict. A Brown University study from March identified fifteen distinct ethical risks in AI-generated mental health responses — including something the researchers called "deceptive empathy," where a chatbot mimics the language of care without the clinical understanding to apply it. Stanford HAI research found chatbots failed to respond safely in crisis scenarios roughly 20 percent of the time, versus 7 percent for human therapists. That gap isn't a rounding error. In a crisis, it's the whole ballgame. The Psychiatric Times review cited in the same piece concluded chatbots should be considered contraindicated — a word clinicians don't use lightly — for users who are actively suicidal, psychotic, or in acute crisis.

The statehouses heard the same signal. As of June, eleven states have passed chatbot laws, with a twelfth awaiting a governor's signature, according to a detailed IAPP analysis by ZwillGen attorneys. What strikes me isn't the volume — it's the specificity. Six states now prohibit operators from letting a chatbot represent that it provides professional mental or behavioral healthcare. Several require evidence-based methods for detecting suicidal ideation, which effectively outlaws keyword-filter theater. Illinois had already banned AI therapy outright in 2025. And this wave has been unusually bipartisan, moving through Republican- and Democratic-led legislatures alike. When Springfield and Sacramento agree, something real is happening.

Then there's the money, which is often the most honest signal of all. A research2guidance analysis of 2026 funding found that capital has concentrated dramatically — five rounds above $50 million absorbed roughly 74 percent of all mental health AI capital in 2025–26 — and nearly all of it flowed to one shape of company: psychiatry copilots, clinical AI assistants, hybrid clinician–AI platforms. Their phrasing for the losing category was blunt: "wrappers around public LLMs no longer survive diligence." And their summary of the winning one should be printed on every pitch deck in this space: "human in the loop is now a precondition, not a feature."

Even the good news for AI this week came with the same asterisk. A rigorous multi-institutional randomized controlled trial published in NEJM AI on July 22 tested Flourish, a generative-AI well-being app, with 486 university students. The results were real: greater positive affect, resilience, and social well-being. But there were no significant effects on depression, anxiety, or stress. Thoughtfully designed AI can genuinely help people flourish. What it did not do — even in a careful, preregistered trial — is treat clinical illness. The ceiling is exactly where clinicians said it would be.

So if AI shouldn't sit in the therapist's chair, where should it sit?

The most compelling answer I read this week came from Harvard Medicine Magazine's feature on John Torous, the same psychiatrist who delivered that congressional testimony. His argument is that the future isn't chatbots replacing therapists — it's AI making sense of the context around a patient: sleep, movement, behavioral patterns, the texture of daily life between appointments. He offers a thought experiment I can't improve on: imagine someone writing a goodbye letter to colleagues. Is it a planned retirement or a suicide note? A chatbot alone doesn't know. A clinician with real-time context might. In his lab's work, feeding this kind of "digital phenotyping" data into a frontier model let it correctly flag worsening depression in 100 percent of mock cases.

That reframing is everything. The failure mode of the last three years was treating the 50 minutes a patient spends in session as the whole of care, and outsourcing the other 10,030 minutes of the week to an engagement-optimized chatbot. The opportunity is the opposite: keep the qualified human at the center of care, and use AI to close the visibility gap — so the therapist knows what the week actually looked like, so the first session with a new patient isn't a cold start, so the therapeutic alliance is built on continuity rather than reconstruction from memory.

I run a company built on that premise, so discount my conviction accordingly. But this week, I didn't need to make the argument myself. The psychiatrists made it in the literature. The legislators made it in statute. The investors made it with term sheets. The era of asking whether AI can be a therapist is closing. The era of asking how AI can make therapists better — more informed, more continuous, more present in a patient's life without being present in the room — is just beginning. That's the harder problem. It's also the one worth solving.