My private practice AI policy

Machine distillation of therapy can short-circuit the most meaningful parts of the process.

I wrote an AI policy for my private psychiatric practice.

In other domains, I evaluate AI use cases with a simple principle: does this tool expand or contract my humanity? This principle is personally useful, but it doesn’t communicate boundaries to my clients.

I started with two questions: (1) how will I use LLMs in decision-making, and (2) when will I use an ambient AI scribe?

The first question was easy to answer: I don’t discuss patient cases with an LLM, and I don’t directly use generative models in decision-making. This isn’t a change to my practice, and so it doesn’t impose a burden. Importantly, the hardest part of my work is the real-time application of “soft skills.” That’s not something I can outsource to AI.

(On the other hand, LLMs are invaluable for finding scientific literature. But I don’t outsource critical thinking; I personally review the relevant papers.)

The second question required more thought, but I ultimately decided not to use an AI scribe.

AI scribes make sense in most medical settings. When electronic health records were introduced, the tradition of minimalist note-taking slowly melted away. Bloated records became expected, and incentives pushed clinicians to spend more time with computers and less time with patients. AI scribes reduce late-night charting, which is a win for clinicians. This doesn’t necessarily mean that patients get more time with their doctor, but some of my colleagues are investing their time savings back into patient care.

My concerns are more subtle. An AI scribe is a third party to conversations. This changes the discussion, and that matters in psychotherapy—especially when someone is opening up about trauma, substance use, infidelity, religious doubts, non-disclosure agreements, or suicidality. Even a sandboxed, local LLM encourages self-censorship. That changes the therapeutic relationship in a way we haven’t adequately studied.

More importantly, AI scribes misrepresent therapy in a way that can’t be corrected with greater context. Therapy is not an information problem. (The bitrate of knowledge transfer is trivial.) Most of the automated summary is irrelevant, and the parts that matter most may be completely missed. How did the relationship progress? Which parts of the conversation had the most emotional resonance? What subtle hypotheses do the client and clinician refine? Machine distillation of therapy can short-circuit the most meaningful parts of the process. That’s not an idle concern. My school of therapy prioritizes clear explanations and rapid experimentation. But it works best when the client does most of the discovery. The journey is more important than the destination.

This approach won’t work for every psychiatrist. It’s my attempt to mindfully use technology in a way that matches my values and practice style.