The Future of Mental Health Care Needs AI in the Right Place, Not Every Place

Of every question I get asked about healthcare AI, the one I am most careful with is some version of can AI be someone's therapist. I understand exactly why the question comes up. Access to licensed mental health professionals is genuinely scarce in large parts of the world, wait times for a first appointment can stretch for months in places I have worked, and a chat interface that is available at two in the morning feels, on the surface, like it is solving an access problem that badly needs solving. But my answer has not moved and will not move, because the honest answer is no. A licensed therapist or psychiatrist brings clinical training, the ability to recognize risk that does not announce itself in words, legal and ethical accountability, and a therapeutic relationship built on continuity and trust that a model cannot replicate regardless of how fluent its responses sound. Presenting an AI system as a substitute for that relationship is not an innovation, it is a category error with real consequences for people at their most vulnerable.

I say this as someone whose company builds AI for healthcare, not as someone skeptical of the technology in general. It is precisely because I believe in what AI can do well that I am strict about what it should not be asked to do. Mental health is the domain in this industry where the cost of getting the boundary wrong is highest, and that is exactly why the boundary has to be drawn clearly rather than left to marketing language to blur.

None of that means AI has no role in mental health care, and dismissing it entirely would be its own kind of disservice to people who need better access than the current system provides. The role I have seen genuinely help sits around the edges of clinical care rather than inside it. On the access side, a well-built triage tool can help someone figure out what kind of support they actually need and how urgent it is, distinguishing between someone who would benefit from a structured self-help resource, someone who should see a counselor within the week, and someone whose situation requires immediate escalation to a crisis line or emergency service. That routing function, done conservatively and with a strong bias toward escalating rather than under-reacting, extends the reach of a scarce professional workforce without pretending to replace it.

On the administrative side, AI can meaningfully reduce the burden that keeps therapists from seeing more patients, drafting session notes for clinician review, handling scheduling and intake paperwork, and surfacing relevant history so a clinician walks into a session already oriented. And between sessions, structured tools like guided journaling prompts, mood tracking, or reminders tied to a treatment plan a clinician has actually set can support what a patient and their therapist are already working on together. The common thread across all of these is that AI is operating in service of a licensed professional's care plan, not operating instead of one.

The line I hold firmly is that AI should never be positioned, marketed, or quietly allowed to function as a replacement for diagnosis, crisis intervention, or ongoing therapeutic treatment. A person in crisis needs a trained human being who can exercise judgment that no model I have evaluated can reliably exercise, judgment about risk, about what is being left unsaid, about when a pattern of responses signals something that requires immediate human intervention rather than another automated reply. Any product in this space that does not build hard escalation paths to real crisis resources, and does not make unmistakably clear to users that they are talking to a tool and not a clinician, is, in my view, not ready to be in front of vulnerable people regardless of how sophisticated its underlying model is.

This is also why I am uncomfortable with how casually some products in this space use the word therapy. Between-session support, triage, and administrative relief are legitimate and valuable categories of AI assistance. They are not therapy, and calling them that, even implicitly through product framing, sets an expectation that can cause real harm when a person leans on the tool in a moment that actually calls for a licensed professional.

Mental health is a category where we have chosen to move more slowly and more conservatively than the market pressure around us would suggest, because I would rather build nothing than build something that fails someone at the moment they need help most. The work we have done in this space has focused narrowly on the access and administrative layer: helping mental health providers triage incoming patients so the more urgent cases are seen sooner, reducing the documentation load that keeps therapists from carrying larger caseloads, and building clear, prominent pathways to human crisis support into every tool we ship rather than treating that as an afterthought. The goal is a system where more people reach a licensed professional faster, not a system where fewer people need to, because that second goal is not one AI is equipped to deliver responsibly today, and I do not think it should pretend otherwise.