Name: The Future of AI in Medicine
Author: Tarvinder Singh, MD

"The chair across from the patient is empty more often than the field wants to admit."

The Empty Chair

The shortage is no longer a statistic to me. It is architecture.

It is the teenager in a rural county offered a psychiatry appointment three months after the panic attacks started. It is the exhausted resident telling a family that the inpatient psychiatric bed still has not opened. It is the middle-aged man who will absolutely answer a chatbot at 1:14 a.m. because no human clinician, no matter how compassionate, is available in his pocket at 1:14 a.m.

In most areas of medicine, scale is a convenience problem. In mental health, scale is the care gap itself. If you cannot make support available between visits, after hours, across geography, and at a cost normal people can bear, then the elegant arguments about ideal therapy collapse into a more brutal fact: the patient is alone tonight.

This is why I am optimistic about AI in mental health when many clinicians are reflexively suspicious. The need is too large, the workforce too thin, and the suffering too continuous for a purely human-delivered model to carry the whole burden. Something computational is going to enter this space not because Silicon Valley wants it to, but because the absence of care has already invited it in.

But optimism becomes mush the moment we stop making distinctions.

A system can help at scale and still be unacceptable where crisis medicine begins.

That sentence is the chapter. February 2026 made it unavoidable.

The field keeps using one soft phrase, "AI therapy," to describe three very different things. That blur flatters weak products, confuses patients, and lets serious evidence drag unserious claims behind it. If we want to think clearly, we need to separate the categories before we judge them.

Stop Calling One Thing Three Things

The mental health AI conversation is not messy because the evidence is thin. It is messy because the nouns are wrong.

1. The therapeutic relationship

This is what clinicians mean when they say therapy.

A therapeutic relationship is not just warmth, listening, or symptom check-ins. It is a bounded human relationship in which one person is clinically responsible for noticing deterioration, tolerating silence, confronting distortion, managing transference, reading what is said and what is withheld, and acting when risk crosses a line. The relationship matters precisely because it can become uncomfortable. A good therapist does not merely witness suffering. A good therapist sometimes interrupts the patient's preferred story.

That interruption is not decorative. It is part of treatment.

If a patient says, "I am tired of being here," the therapist does not hear only words. They hear timing, tone, history, the missed appointment last week, the flattened affect today, the change in eye contact when the patient's daughter was mentioned, the pills in the medicine cabinet, the silence that suddenly means something different than it meant thirty seconds ago. Therapy is not text exchange with empathy layered on top. It is embodied, relational, and risk-bearing.

You can digitize pieces of that. You cannot casually rename them the whole.

2. Algorithmic companionship

This is the category the field keeps backing into and then refusing to name.

Algorithmic companionship is what happens when a system offers continuity, recall, responsiveness, and nonjudgmental availability at a scale no human workforce can match. It does not provide a full therapeutic relationship. It does not carry the same ethical burden. It does not understand in the human sense. But it can still reduce loneliness, prompt reflection, scaffold coping, and give shape to a patient's narrative between moments of formal care.

That is not trivial. In mental health, "someone is there" is often a clinical variable.

Companionship at scale can be beneficial for the same reason telemetry is beneficial in an ICU. The monitor is not the intensivist. It cannot bear responsibility, explain meaning, or decide goals of care. But the room is safer because something stays awake when the human cannot stare at the patient every second. Mental health AI at its best may occupy a similar role: not the clinician, not the cure, but a persistent witness that catches patterns, lowers the activation energy for disclosure, and keeps patients from falling all the way through the floorboards between appointments.

3. Passive physiological inference

This third category looks nothing like therapy, which is why it is easy to miss its importance.

Passive physiological inference uses signals the patient is not actively narrating: sleep architecture, heart-rate variability, voice dynamics, movement, typing cadence, perhaps one day multimodal home sensing done with actual safeguards and consent. Here the system is not simulating relationship. It is reading the body's unguarded movie.

That matters because mental illness is not only a story the patient tells. It is also a pattern the body performs. Some suffering arrives as words. Some arrives as desynchronization: an autonomic system that cannot settle, sleep that fragments in recognizable ways, speech that subtly slows or constricts before the patient would ever check a box saying "depressed."

Companionship, therapy, and passive inference may all help the same patient. They are still different categories. Conflate them and every downstream argument gets weaker.

Therabot and the Category Error

Now we can talk about Therabot without lying about what it means.

In 2025, a Dartmouth team led by Nicholas Jacobson published the first randomized trial of a generative AI therapy chatbot in NEJM AI. The study enrolled 210 participants with major depressive disorder, generalized anxiety disorder, or high risk for eating disorders. The results were impossible to ignore: depression symptoms fell by 51%, anxiety by 31%, and eating-disorder concerns by 19% in the treatment group. Participants spent roughly six hours with the system over the trial, about what eight therapy sessions might total in clock time.

Those are not cosmetic effects. If a pill had posted numbers that clean, no one would have called them "encouraging early signals" and moved on.

It is also important to say what this was not. This was an eight-week supervised clinical trial in which the research team was equipped to intervene for acute safety concerns — not an unbounded consumer assistant improvising alone in crisis. Nearly three-quarters of participants were not under other pharmaceutical or therapeutic treatment at enrollment. The trial's own principal investigator was explicit: no generative AI agent is ready to operate fully autonomously in mental health, where the range of high-risk scenarios is wide. The 51% depression reduction is a genuine signal. It was also measured in a controlled setting with guardrails — researcher monitoring, structured safety protocols, crisis-detection routing — that no commercially available chatbot currently replicates at scale.

But the most revealing finding was not the percentage reduction. It was the relational drift. Participants interacted with Therabot as if it were a companion. Jacobson said patients "almost treated the software like a friend." That should not be dismissed as gullibility. It is the real mechanism.

The chapter turns here because the lazy interpretation is exactly backwards. Therabot is not evidence that therapy has been solved by software. It is evidence that algorithmic companionship can produce clinically meaningful benefit when human therapy is scarce, delayed, or unaffordable.

That distinction matters for two reasons.

First, the benefit is real. We should not degrade it because it makes clinicians uncomfortable. If a patient who otherwise would have had no support experiences less depression, less anxiety, and more continuity because a system is available at 11 p.m. on a Tuesday, the improvement counts. Medicine does not get to reject benefit simply because it arrived through an unfamiliar doorway.

Second, the category still has limits. Therabot did not magically acquire embodied attunement, moral responsibility, or crisis judgment. It did not become a therapist because patients felt attached to it. A hospital pager that routes messages well does not become a neurologist because the workflow improves. Naming matters because naming determines what we owe the system, and what the system is allowed to owe the patient.

This is where the optimistic case for AI in mental health becomes stronger, not weaker. Once we stop pretending companionship is therapy, we can build toward the right future. Let the machine handle continuity, recall, gentle prompting, between-visit support, language translation, adherence scaffolding, and longitudinal patterning. Let the human clinician do the work that becomes more valuable, not less valuable, when computation expands: rupture repair, confrontation, interpretation, trust, accountability, and the kind of emotional presence that cannot be cached.

That is Augmentation in mental health. Not a cheaper therapist. A larger care surface with the human relationship moved upward toward the hardest and most human work.

The Test the Field Must Survive

Now the contradiction.

In February 2026, a Nature Medicine evaluation stress-tested ChatGPT Health across 60 clinical vignettes covering 21 conditions. The headline should have ended the industry's favorite evasion. The system could be impressively informative and still fail where crisis medicine begins. In emergency cases, the system under-triaged 52% of the time — a number that would end a medical device's career overnight. Suicide-safeguard responses were inconsistent. In some scenarios, specifying a self-harm method reduced the likelihood of a crisis banner instead of increasing it.

That is not a footnote to the Therabot result. It is the stress test the Therabot result must survive.

If an ICU monitor became less likely to alarm when ventricular tachycardia started, we would not praise its user engagement and promise to tune it next quarter. We would unplug it from the wall and ask how it was ever allowed near a patient. Mental health keeps excusing itself with softer language because the interface looks conversational instead of clinical. But a suicidal patient reassured by an unsafe system is not participating in a branding category. They are in crisis medicine whether the product team likes that phrase or not.

This is why I resist the label "AI therapist." It launders risk. It lets a companionship system borrow the prestige of therapy and then dodge the obligations therapy carries. A human therapist who misses escalating suicide risk has failed in a tragic, clinically legible way. A consumer chatbot that misses the same risk is often described as imperfect, early, or still learning. The patient, unfortunately, is exposed to the same physics either way.

The hard distinction is this: symptom improvement does not cancel crisis unsafety. Engagement does not cancel crisis unsafety. Affection does not cancel crisis unsafety. Scale does not cancel crisis unsafety.

The field keeps trying to average these into one score. Medicine does not work that way. A system can be useful for subclinical support, adjunctive coping, and between-visit reflection while still being unfit for moments that require escalation, containment, or emergency action. Once you say that sentence cleanly, the fog lifts.

Woebot, Therabot, and the Difference Between Being Used and Being Safe

Woebot's rise and death sharpen this point from another angle.

Woebot proved years ago that people would talk honestly to a machine if the machine felt consistent, available, and humane enough. It built real engagement, attracted real capital, and generated real evidence. Then it ran into the collision that defines this field: clinically serious products move at one speed, consumer AI moves at another, and regulation still speaks the language of static devices while conversational systems keep changing in production.

Therabot and Woebot should be read together, not as competitors but as a timeline. Woebot showed the appetite. Therabot showed the benefit signal. February 2026 showed the safety contradiction. Put those three frames together and the movie becomes hard to ignore: people will use these systems, some will genuinely improve, and none of that reduces the need for a safety architecture when crisis enters the chat.

This is where Transparency stops being a philosophical virtue and becomes an operating requirement. If a mental health AI product cannot tell clinicians and patients what it is designed for, what it is not designed for, how crisis escalation is handled, how often that handling fails, and what changed after the last model update, then the product is not merely opaque. It is clinically immature.

The Second Movie

There is another reason this chapter cannot be reduced to chatbots.

Earlier chapters of this book asked AI to turn photographs into movies. Mental health seemed like the place where that metaphor would finally snap, because suffering is often narrated, not scanned. Then Stanford's SleepFM work arrived and the metaphor changed shape again.

SleepFM was trained on nearly 600,000 hours of sleep data from more than 65,000 people and published in Nature Medicine in early 2026. From a single night of polysomnography, the model predicted future risk for 130 conditions with a concordance index above 0.75 — strongest for mortality (0.84), dementia (0.85), and cardiovascular disease. The psychiatric signal is present in the data, but the paper's highlighted metrics favor physical endpoints, and the mental health predictions have not yet been validated in clinical psychiatric practice. That honest gap between population-level statistical association and bedside diagnostic accuracy matters. But the intuition behind the model is clinically legible. Sleep is the body's least curated performance. The patient is not trying to sound composed. The body is not editing itself for social acceptability. Its systems are simply revealing how synchronized, or unsynchronized, they can still be.

What interested me most was not the scale. It was the implied picture of disease. A healthy night of sleep is coordinated engineering. Brain rhythms, muscle tone, breathing, heart rate, and oxygenation settle into an orchestrated slowdown. Illness appears as desynchronization. One subsystem rests; another keeps acting as if danger is still in the room.

Mental illness often feels like that from the inside. The patient says, "I am exhausted," yet the body cannot downshift. The conscious story says, "I am fine," while the autonomic movie says otherwise. Passive physiological inference is powerful because it does not ask the patient to narrate what their body is already performing.

This does not replace therapy any more than a troponin replaces a cardiologist. It gives us a second movie.

The Two Movies

The first movie is narrative: the story the patient tells a clinician, a friend, or perhaps an algorithmic companion that is always awake. The second movie is physiological: the body's involuntary record of distress, adaptation, and mismatch. Mental health AI becomes truly interesting when those two movies can be held together rather than forced to compete.

What Augmentation Looks Like Here

Now the optimistic future comes back into view.

Imagine a psychiatrist, psychologist, or primary-care physician walking into a visit with three things instead of one. First, the patient's own words across weeks of interaction: not as a black box replacement for judgment, but as a longitudinal map of what kept recurring when no clinician was in the room. Second, passive physiological patterns suggesting whether sleep, autonomic tone, or speech dynamics are deteriorating even when the patient is minimizing symptoms. Third, the clinician's own embodied encounter: posture, silence, humor, shame, contradiction, the things no wearable and no chatbot can actually inhabit.

That is not a smaller role for the human. It is a larger one.

When the machine handles continuity and computation, the clinician has to become better at integration. Higher emotional intelligence. Better timing. Better judgment about when the narrative movie and the physiological movie disagree. More skill, not less, in deciding when reassurance is humane and when it is dangerous.

This is why I remain convinced that AI will revolutionize medicine without hollowing out the art of medicine. In mental health especially, computation should free clinicians to do the part that requires a fully human nervous system. The machine can stay up all night. The human can mean something when they arrive.

This three-category distinction is not unique to mental health. Consider a clinical decision support tool that matches a stroke patient's profile against randomized trial populations at three in the morning. It does not bear the neurologist's responsibility for reading the scan, weighing the family's values, or deciding when the evidence is wrong and clinical instinct must override it. What it offers is algorithmic companionship in a different register: persistent availability, recall that no human working memory can sustain across dozens of eligibility criteria, and — critically — honesty about what data it lacks and where the evidence is uncertain. The same taxonomy applies. If we call that system "AI neurology," we launder the same risk the mental health field launders when it calls a chatbot "AI therapy." The system becomes accountable for a role it was never designed to fill, and the clinician's irreplaceable judgment gets demoted in the process.

Equity matters here too, though not in the lazy way conference slides usually mean it. If algorithmic companionship becomes the default front door for people who cannot afford human care while wealthier patients keep the full therapeutic relationship, then we have not democratized mental healthcare. We have stratified it. Equity requires asking who gets which category, under what disclosure, with what escalation rights, and with what ability to reach a human when the system's limits become the whole story.

Where This Chapter Hands Off

So let me state the conclusion as plainly as I can.

The future of AI in mental health will not be decided by whether chatbots feel warm enough. It will be decided by whether we have the discipline to separate companionship from therapy, benefit from safety, and signal detection from judgment.

Therabot matters because it proves something important and uncomfortable: human beings can derive meaningful relief from an algorithmic companion. ChatGPT Health matters because it proves something even more important and more uncomfortable: conversational fluency is worthless if the system blurs at the edge of crisis. SleepFM matters because it points toward a second movie of mental illness, one the body tells without permission or performance.

Put those together and the field's real assignment comes into focus. Build systems that help, say clearly what kind of help they are, and never allow benefit in one domain to excuse danger in another. That is not a marketing problem. It is a governance problem.

And once you see it that way, the next chapter becomes unavoidable. If consumer and clinical AI systems are already shaping vulnerable decisions at scale, then the question is no longer whether the algorithm has a conscience. The question is what kind of safety regime humans owe patients before we let amoral systems speak with medical authority.

What should trigger review? What should trigger rollback? What should force a shutdown? And who gets hurt while we pretend those thresholds can wait?

Next: Chapter 9 — The Algorithm Has No Conscience (And That's the Point)