Psychiatrist — tasks, one by one

Assessment interview and mental state examination

The examination is made of observation in the room: affect, pace of speech, eye contact, what the person avoids saying, how they respond to a question asked a second way, and what a relative says that the patient did not. Questionnaires and chatbots can collect a structured history before the appointment, but people with psychosis, mania, severe depression or cognitive impairment are often the least able to fill one in reliably, and a history that sounds coherent can still be wrong. Recording and transcription help with what was said; they do not capture what was seen, and many patients are reluctant to be recorded at all.

This judgement covers the examination, not the intake around it; pre-visit questionnaires and triage tools could take over the first contact and shorten appointments without changing who examines. Video consultation already changes where the interview happens. A deployment at scale where a system's own assessment, without a clinician's interview, is used to decide care would move this.

Making the diagnosis

Part of diagnosis is reading: long records from earlier admissions, letters from other services, scored questionnaires, and tests that rule out a physical cause. Language models can summarise a thick file, line up symptoms against diagnostic criteria and list differentials worth considering, which saves time and can catch something missed. But psychiatric diagnosis has no single test to anchor it; it rests on the clinician's reading of the interview, and licensing rules make the doctor answerable for the label, which follows a person into insurance, employment and legal proceedings. So the machine shortens the preparation and the doctor still decides.

Faster preparation is not fewer psychiatrists; it may instead let the same doctor see more people, or be absorbed by rising referrals. Summaries of records can omit or invent details, and the cost of that error falls on the doctor who signed. A validated diagnostic tool that a regulator allows to be used without a psychiatrist's confirmation would move this towards automating.

Prescribing and monitoring medication

Electronic prescribing already checks interactions, doses and allergies, and monitoring schedules for drugs that need regular blood tests can be tracked by software. Newer tools can draft a medication review or flag a pattern in results. What stays with the doctor is the choice itself: weighing a side effect the patient fears against a relapse they cannot afford, judging whether a person will take the drug, and taking legal responsibility for a prescription of medicines that are often sedating, sometimes dangerous in overdose and in some cases controlled. Prescribing authority is tied to a licence, and that tie is what keeps the result with a person.

The protection here comes from prescribing rules, which are a policy choice and can widen to other professions as well as to software; wider prescribing rights for nurses and pharmacists may change psychiatrists' share of this work sooner than any machine. A rule letting software adjust or renew psychiatric medication without a prescriber's sign-off would move this.

Psychotherapy

Chatbots that hold a supportive conversation or walk a person through structured exercises exist and are used for milder distress. The patients a psychiatrist treats are usually the more severe and complex ones, where the therapeutic relationship carries much of the effect and where a model that agrees too readily, misses a change in risk or reinforces a delusional belief does harm. Rules on who may provide therapy, and professional bodies' caution about unsupervised chatbots, slow adoption further.

This says nothing about how much therapy psychiatrists still do; in many health systems it has already moved to psychologists and therapists, and self-guided digital programmes may take the milder cases first. A controlled trial in which a system delivers therapy to people with severe illness, compared with a clinician rather than with no treatment, followed by adoption in services, would move this.

Risk assessment and involuntary treatment

Risk tools that score a person from their history have long existed and predict poorly for an individual, which is why risk is judged in conversation and in context. The decision to detain or treat without consent is the most constrained act in the job: mental health laws name who may make it, require a medical examination, and open it to review by a tribunal or court. That accountability cannot be handed to software without changing the law, and the cost of an error in either direction, a death or a wrongful detention, makes services unwilling to try.

The protection is legal and could change, but it binds the decision, not the paperwork: forms, tribunal reports and records around a detention can still be drafted by machines. Monitoring tools that flag risk from records or phone data could shift when a psychiatrist is called in. A mental health law that lets an automated assessment trigger or extend a detention would move this.

Notes, letters and reports

This is the most text-heavy part of the work and the one language models fit best: ambient tools draft notes from a consultation, and a model can turn a long record into a first draft of a letter or report. The doctor still reads, corrects and signs, because the document carries a diagnosis and an opinion that someone else will act on. Adoption in psychiatry is slower than elsewhere in medicine because the conversations are among the most sensitive a health system records, so consent, storage and who can read the transcript are harder questions.

Time saved on writing is not reduced need for psychiatrists; in services with long waiting lists it is likely to become more appointments rather than fewer posts. Reports for courts and tribunals are opinions given under the doctor's name, so drafting help does not change who must give them. A large deployment of ambient documentation in psychiatric services with measured effects on clinician time or staffing would sharpen this judgement.