AI in Psychiatry — What It Does and Doesn't Do

Home > AI in Psychiatry

I both build these systems and study them. That combination is unusual and it shapes what I think.

What I've actually published

MIND: Empowering Mental Health Clinicians with Multimodal Data Insights through a Narrative Dashboard (2026) — a peer-reviewed, co-designed LLM-powered dashboard that presents patient-generated data alongside clinical notes as narrative text rather than raw charts. In a user study with 16 clinicians, MIND significantly outperformed baseline methods at surfacing clinically relevant insights (p<.001) and supporting decision-making (p=.004).

That's a clinician-facing tool. Note what it does: it helps a psychiatrist see the data. It does not make the decision.

Digital psychiatry oversight — published work in JAMA Psychiatry on the regulatory gaps around unregulated telehealth and digital mental health platforms.

The Sultan Lab at Columbia uses natural language processing on electronic health record data to identify treatment gaps — current work examines nicotine vaping in patients with comorbid psychiatric conditions through clinical notes.

Three categories that keep getting conflated

1. Clinician-facing tools. Documentation, note generation, decision support, literature synthesis. This is where the evidence is strongest and the risk is lowest, because a licensed clinician remains between the model and the patient. It's also where my own work sits.

2. Structured therapeutic applications. Purpose-built programs delivering a specific protocol — CBT for insomnia, exposure hierarchies — with clinical design and, ideally, oversight. Some randomized evidence exists for specific products. Not the same as a chatbot.

3. General-purpose chatbots used as therapy. Millions of people are doing this. The evidence base is thin and the documented failure modes are real: inconsistent handling of suicidal ideation, sycophancy that reinforces distorted beliefs rather than challenging them, no capacity to escalate, no continuity, and no accountability when it goes wrong.

Conflating these three is how both the hype and the panic get generated.

What I tell patients who ask

If you use an AI chatbot between sessions, tell me. I'm not going to lecture you about it. But I want to know what it's telling you, because I've seen it agree with things I'd have pushed back on. That agreement feels supportive and is sometimes the opposite of helpful.

It cannot assess risk. A model has no access to your history, your baseline, or what you look like when you're not okay. Risk assessment is the part of psychiatry that most requires knowing a person over time.

It is very good at some things. Explaining a diagnosis at 2am. Drafting what you want to say to your boss. Organizing your thoughts before an appointment. Understanding a medication you were prescribed. These are real uses and I don't discourage them.

Privacy is not what you think it is. A general-purpose chatbot conversation is not protected health information. It is not covered by HIPAA, and depending on the product it may be retained, reviewed, or used for training.

Where I think this goes

The useful question isn't whether AI replaces psychiatrists — it doesn't and won't, for reasons that have to do with risk, accountability, and the fact that psychiatric assessment is substantially about what a person doesn't say.

The useful question is whether it fixes the part of psychiatry that's genuinely broken: that the average outpatient visit has compressed to fifteen minutes largely because of documentation burden and reimbursement structure. If these tools return clinical time to clinical work, that's a real gain. If they're used to compress visits further, that's a loss disguised as efficiency.

That's the design question I care about, and it's why I build clinician-facing tools rather than patient-facing ones.

Related

Medication management · ADHD telehealth apps vs. in-person evaluation · How to choose a psychiatrist · Publications · AI in psychiatry: research, digital therapeutics, and responsible innovation · Sigmund

Frequently Asked Questions

Can AI replace a therapist or psychiatrist?

No. Current systems cannot assess risk, cannot escalate in a crisis, have no longitudinal knowledge of a patient, and carry no accountability. Psychiatric assessment depends substantially on history, context, and what a person doesn't say.

Is it safe to use ChatGPT for mental health support?

For some purposes — understanding a diagnosis, organizing your thoughts, drafting a difficult conversation — it's genuinely useful. For risk assessment, crisis support, or as a substitute for treatment, no. Documented failure modes include inconsistent handling of suicidal ideation and agreeing with distorted beliefs rather than challenging them.

Are AI therapy chatbots evidence-based?

It depends entirely on which one. Purpose-built applications delivering a specific protocol have some randomized evidence. General-purpose chatbots used as therapy have very little, and they weren't designed for it.

Do you use AI in your practice?

For clinician-facing tasks — documentation and literature synthesis — with the clinical decisions remaining mine. I co-authored a peer-reviewed evaluation of exactly this kind of tool.

Is my conversation with an AI chatbot private?

Generally not in the way a medical record is. General-purpose chatbot conversations are not protected health information, are not covered by HIPAA, and depending on the product may be retained or reviewed.

Should I tell my psychiatrist I use AI for mental health?

Yes, and you won't get a lecture from me. It's useful to know what it's been telling you — particularly if it has been agreeing with something that warrants challenging.