Ambient AI has taken off quickly in somatic medicine. The principle is simple: a microphone captures the consultation, a model produces the note, the physician saves time. In general practice, gastroenterology, orthopedics, the gain is real and the clinical loss limited.

In mental health, the equation is different. And the main problem isn’t technical. It’s clinical.

The therapeutic setting depends on what isn’t recorded

In psychiatry and psychotherapy, the session rests on a founding asymmetry: the patient agrees to say things they tell no one else, in exchange for a frame where that speech is protected. This frame holds only through a small number of implicit conditions. Two people in a room. No witness. No archive of raw speech. Medical confidentiality covers what the clinician retains, not everything that was said.

Introducing a recording, even “zero-retention,” even “encrypted,” even well-intentioned, changes this architecture. The patient knows a microphone is on. They don’t know what will be processed, by whom, or for how long. Even when it’s explained, part of them stays cautious.

What patients stop saying

What disappears isn’t randomly distributed. It’s systematically the most clinically useful content:

  • Suicidal ideation, especially precise plans or intent
  • Substance use — alcohol, cannabis, non-prescribed benzodiazepines, cocaine, recreational ketamine
  • Acting out — intimate partner violence, sexual assaults experienced or committed, risk behaviors
  • Hallucinatory content experienced as shameful or bizarre
  • Doubts about the therapeutic relationship itself — “I don’t know if this is helping,” “you remind me of my father,” erotic transference
  • Intimate content — sexuality, unprocessed grief, identity, abortions, old family violence

None of this appears in a somatic medicine note. All of this is the raw material of psychological work. Losing this material to a recording device means losing the session.

Studies of recorded consultations (in clinical research, where it’s been done for a long time) show a documented reduction in spontaneous disclosure. Patients self-censor. They don’t say so. They adapt their narrative to an audience they imagine — and they’re not wrong: the microphone is a witness.

The therapeutic alliance isn’t a placebo effect

A classic objection: “the patient gets used to it, the effect disappears after a few sessions.” This isn’t true in the same way for all content. The patient can get used to the microphone for trivia. They don’t get used to it for content that threatens their self-image — and those are precisely the ones we want to hear.

Beyond that, the therapeutic alliance isn’t just subjective. It’s predictive: the quality of the alliance remains the best predictor of therapeutic outcome, across approaches. Anything that weakens it weakens treatment, not just patient comfort.

Some jurisdictions (Switzerland, France, EU) require explicit consent for any recording of a consultation. Consent is necessary, but it doesn’t solve the clinical problem. A patient who consents reluctantly, because refusing feels socially awkward or because “everyone uses this tool now,” consents without being truly free. And formal consent doesn’t rebuild the lost frame.

There’s also a practical question rarely asked: what happens when the patient says “wait, turn it off, I want to tell you something”? In principle, you turn it off. In practice, the conversation that follows is the one that should have happened from the start — and the rest of the session, recorded, is clinically poor by comparison.

Even if the intent is benign, an audio recording of a psychiatric consultation exists, somewhere, at some point. What exists can be:

  • Subpoenaed by a judicial authority in civil, criminal, or disciplinary proceedings
  • Requested by the patient under GDPR / Swiss FADP right of access
  • Leaked in case of provider compromise
  • Used against the clinician in a complaint

None of these hypotheses is theoretical. We discuss this in detail in a dedicated article on the legal risk of recordings. The key point here is that the only sure way not to see a recording become a problem is not to have it.

The alternative: nothing during, everything after

The right compromise for mental health is what many clinicians already practice without thinking about it: nothing is recorded in the session, and the note is dictated after.

Concretely: after the session, while clinical thinking is still fresh, the clinician dictates two to four minutes of structured reflection. A tool turns that dictation into a note. Nothing the patient said is recorded. The therapeutic frame is intact. The work of selection and interpretation — the core of the craft — stays in the clinician’s head.

Is it slower than ambient? Marginally. Two to three minutes per session. It’s also more faithful to what a mental health note should be: a reconstruction, not a transcript. See Why the clinical note is written after the session.

Conclusion

Ambient AI is a useful tool when clinical material is factual and the patient has nothing to hide. In mental health, those conditions rarely apply. The microphone changes the session. And what’s lost in the session isn’t recovered in the note.