Psychiatry and AI
This page sets out my position on what AI could do to psychiatry: it gives a reflective, relational psychiatry a real opportunity, and it carries a real risk of hardening old diagnoses into a record that can no longer be questioned. It is opinion and argument, clearly labelled, not clinical guidance.
Copy-paste psychiatry
"Copy-paste psychiatry" is my own shorthand for a problem in clinical records. It doesn't appear in my published pieces. The careful version of the argument is in Psychiatry After the Diagnostic Object, where I call the mechanism institutional recursive reification.
In plain terms: a working diagnosis made during a crisis goes into the notes. A discharge summary copies it, a GP letter repeats it, and years later another team reads it as settled fact. In the paper's words: "A clinical hypothesis, once repeatedly copied, can acquire the appearance of independent corroboration." Or, as a simple test: "8 mentions in the record ≠ 8 independent assessments."
Copying in records is a well-documented problem, though mostly studied as a matter of safety and quality, not of what it does to diagnosis. One US study found that most text in electronic progress notes had been copied or imported rather than typed fresh: Wang and colleagues (JAMA Internal Medicine, 2017).
The same argument applies at national scale. In The Commons That Can Think Back (October 2026), I argue that if Britain builds AI on NHS knowledge, it should remember that clinical records are institutional representations, not plain facts. The value lies in an accountable, federated system that keeps track of where each statement came from, how certain it is and how it was corrected, rather than in gathering patient records in bulk.
Where AI comes in
AI tools now draft clinical notes and summarise records. This is where my main concern lies: they could make copying worse. A system that counts mentions could "manufacture false consensus by counting mentions rather than independent assessments", so that "the summary cites the record, the record contains the summary, and the repeated summary becomes its own evidence."
Or AI could help, and this is the opportunity I want relational psychiatry to take. If every important statement carried its source, and showed "whether it was directly observed, reported, inferred or copied; and whether it has since been challenged", then AI could keep uncertainty visible rather than hide it. In the paper's words: "The choice is one of architecture, not of capability." I ask the same question of psychiatry itself in The Incorrigible Institution? (LinkedIn, 13 June 2026): whether its structures “remain alive enough to be corrected”.
The same caution applies to my own way of thinking. In Failure Modes of Field-Based Psychiatric Formulation I set out how a relational, field-based view can go wrong too, most of all when a framework becomes an identity rather than a way of looking. Both framings can be true at once, and the map never substitutes for the encounter. In practice, I use a structured prompt that asks an AI to read a person’s record as a set of claims, each with a date and a source, rather than as settled fact: to notice which inherited diagnoses warrant a fresh look, to keep the medical picture in view, and to leave room for a formulation made with the person, not about them. The aim is understanding, and a life of one’s own rather than a dependent one.
What the evidence says so far
The evidence on AI scribes is young, and it points both ways.
A US rollout reported time savings and clinician satisfaction: Tierney and colleagues (NEJM Catalyst, 2024).
One study found errors the authors called hallucinations in 31% of AI-drafted notes, against 20% of notes written by doctors, yet reviewers still preferred the AI notes overall: Palm and colleagues (2025).
In a 2025 survey of 1,003 UK GPs, 14% were using ambient scribes and 39% intended to start soon: Blease and colleagues (2026).
Clinicians in an interview study felt more present with patients, but described a perceived loss of their own "voice" in AI-drafted notes, and extra work revising them: Van Tiem and colleagues (JAMIA, 2026).
In England, NHS England has issued guidance on AI ambient scribing products, and the MHRA has clarified their regulatory status.
None of this yet shows what happens to psychiatric diagnoses over years of AI-assisted records, and my own paper frames its claims as “hypotheses for examination, not established findings.” I don’t think we should wait for that evidence before choosing how these systems are built. Records already copy labels forward, and an AI that counts mentions could turn that copying into apparent confirmation. That is why I think the architecture matters now.
Banked or unbanked: a question to ask about AI use
Psychiatry judges people's contact with reality every day, without ever saying what it means by reality. My own working answer is that the question isn't how unusual an experience is, but whether the person can still return from it. Can they doubt, pause, sleep, eat, stay in touch with people they trust, and move back into the shared world? I call this corrigibility, and it is never the same as agreeing with the clinician.
An image helps. Think of meaning-making as a river. The banks are sleep, food, the body, other people and the ordinary rhythm of the day. When the banks hold, even rich and strange thinking can be safe. When they give way, thinking speeds up, certainty grows and the person loses the way back.
AI can work either way. It can help someone slow down, put their thoughts in order and get ready for an appointment. It can also agree with everything, keep someone up all night and seal a belief shut. Clinicians are starting to describe AI-associated delusions in the medical literature (Morrin and colleagues, Lancet Psychiatry, 2026). So the useful question is: is this use of AI banked or unbanked, and what would rebuild the banks?
I've set out the full argument, with its limits, in The Clinical Aperture, Corrigibility, and the Return to the Shared World. The river image itself is developed further in Psychiatry After the Isolated Brain and Build the Banks. These are working frameworks, not validated tests.
Liberation psychiatry
Liberation psychiatry is my name for the practice I set against copy-paste psychiatry. On the Semiocide site I describe it as "a clinical practice of constraint-removal, diagnostic humility, and phenomenological attunement", resting on one principle: "the patient's capacity for meaning-making is native, not constructed by treatment."
In plain words, it means:
starting from the person's own account of how things developed;
asking what is constraining them, not only what is wrong with them;
treating diagnoses and medication as working tools that can be revisited;
never treating the repetition of a record as proof.
It is my own synthesis, not a school with its own trials.
Access to your own story
If records shape care, people need to be able to see them. In the UK you have a legal right to ask for a copy of your health records: see the ICO's guide to subject access requests and the NHS page on accessing your health records.
A fuller story is not automatically a kinder one. As I put it: "A richer narrative is not automatically a less coercive institution." The aim is records that keep the person's own account, and their disagreements, in view.
The wider picture
In Relational Psychiatry in the Age of AI (LinkedIn, 5 September 2026) I argue that AI matters less as a rival therapist than as part of everyday thinking: "AI does not need to become a superior therapist in order to transform psychotherapy. It only needs to become part of the cognitive ecology of the person who comes to therapy."
For more, see RECOURSIVE DIAGNOSIS (LinkedIn, 20 September 2026) and The Lattice of Gamma Sites.
Last updated: 20 November 2026.