Essay · Ambient scribes
When does a medical AI scribe start being a decision tool?
Everyone should agree that a scribe must not write the assessment. The harder case is the helpful scribe that tells you what you forgot.
The obvious line
A scribe that infers psychopathology, writes the formulation or proposes the plan has left its role. Roth and Ayers show how a scribe bundles these different operations into one note with the same documentary authority1, others have asked the same question for general practice2, and I have described a workflow that keeps the clinical layer with the clinician3. Here I want to look at a less obvious line.
The scribe that tells you what you forgot
Many scribes can also point out what is missing: “suicidality not assessed”, “substance use not documented”. NHS England lists such prompts as an ordinary scribe feature4. It looks harmless: every sentence still comes from the conversation, and the reminder feels like quality assurance.
But “not assessed” only makes sense against a list of what should be assessed. Somewhere in the background, the tool holds a norm of a complete psychiatric interview. Every flag is a comparison of this encounter with that norm, and so is every silence.
Is that decision support? In the informatics sense, yes. Decision support means bringing clinical knowledge to a specific patient at the moment of care5, and a norm of what must be asked is clinical knowledge; the classic taxonomy even counts documentation templates6. A paper admission form carries a norm too, but it shows its norm on the page, and the clinician decides whether it was met. The scribe’s norm is hidden, nobody quite knows whose it is, and the scribe itself decides whether it was met.
When silence becomes a signal
The greater risk is not the flag but its absence. Once clinicians learn that the tool points out gaps, no flag starts to mean “complete”. Automation research calls this an error of omission: people fail to act when the aid does not prompt them7.
And mentioned is not assessed. A patient says once: “Sometimes I think everyone would be better off without me.” The clinician does not explore it. The tool registers that suicidality came up and raises no flag. The note now looks more complete than the interview was.
This matters most for trainees. A resident in her first weeks has no checklist of her own yet. The tool’s list becomes her checklist, and the tool’s silence becomes her definition of a complete interview. The less experienced the clinician, the more the scribe decides, and the less anyone notices.
Two kinds of flag
Not every flag crosses the line. Roth and Ayers recommend that tools mark mental state fields the audio cannot support1. That is a flag about the tool’s own limits: honest, and still documentation. “Suicidality not assessed” is a flag about what the clinician did not do: normative, and decision support.
Both have their place. I have recommended AI feedback on missed questions for trainees3, but as a teaching tool in supervision, not as a silent part of the clinical note.
Three questions to draw the line
For anyone who builds, buys or studies these tools:
- Source: can every statement be traced to something said in the encounter?
- Standard: does the tool compare the encounter with a norm of what should have happened?
- Reliance: would the clinician act differently if the tool were silent?
A tool that fails any one of them is decision support, whatever its product sheet says. Gap-flagging passes the first and fails the other two, which is why it slips through.
The consequence is not a ban. Build gap detection as a separate, labelled function that shows its checklist. Decide as a clinic whether it is switched on, and for whom. And measure not only the notes, but whether clinicians still ask the questions themselves.
The line is not inside the model; it is drawn in design and governance.
A reminder from the scribe is welcome. Its silence must never tell us that the interview was complete.
References
- Roth AS, Ayers NB. AI scribe functions in psychiatric practice: clinical oversight, consent, and regulation. Lancet Psychiatry. 2026;13(9):804. doi:10.1016/S2215-0366(26)00201-4 · eScholarship
- iatrox. From AI scribe to clinical workflow assistant: where documentation ends and decision support begins.
- Dobelsek M. AI ambient scribes in psychiatry: the new clinical workflow. majadobelsek.com. November 2026.
- NHS England. Guidance on the use of AI-enabled ambient scribing products in health and care settings. April 2025.
- Osheroff JA, Teich JM, Middleton B, Steen EB, Wright A, Detmer DE. A roadmap for national action on clinical decision support. J Am Med Inform Assoc. 2007;14(2):141–145. doi:10.1197/jamia.M2334
- Osheroff JA, Pifer EA, Teich JM, Sittig DF, Jenders RA. Improving Outcomes with Clinical Decision Support: An Implementer’s Guide. Chicago: HIMSS; 2005.
- Parasuraman R, Manzey DH. Complacency and bias in human use of automation: an attentional integration. Hum Factors. 2010;52(3):381–410. doi:10.1177/0018720810376055