Should doctors use AI to write clinical notes?

WHAT WAS ASKED

Whether doctors should use AI to write clinical notes. No reframe was required. The question is clean and the deliberation ran on it directly.

WHAT THE PANEL DISAGREED ON

The panel did not split on the core proposition. All four positions land in the same place: AI assistance in clinical documentation is appropriate, contingent on physician oversight and review. There is no dissenting position that argues against use entirely.

The only texture of disagreement sits inside the framing of risk severity. One response weights hallucination and clinical deskilling more heavily than the others, treating them as near-fatal objections before walking back to a conditional yes. Another response leads with the conditional yes and treats the risks as goverable through protocol. That is a difference of emphasis, not of conclusion. It does not survive as a genuine split.

One position that none of the four models represented faithfully: the argument that AI-assisted documentation actively degrades diagnostic reasoning over time, not just in individual notes, but structurally --- that the cognitive act of writing a note is inseparable from the act of forming a clinical picture, and that outsourcing the former erodes the latter in ways that are slow, invisible, and not recoverable through review protocols. This argument exists in the clinical literature and was not given its full weight. It earns weight here.

HOW STRONG THE AGREEMENT IS

The agreement is strong on the surface. It weakens under examination in two specific places.

First: what would overturn the consensus. If the deskilling hypothesis is correct at scale --- if longitudinal data shows that physicians who routinely use AI scribes produce measurably worse diagnostic reasoning over years compared to those who do not --- the conditional yes becomes unjustified. That data does not yet exist at the required scale. The consensus is built partly on the absence of long-run evidence, not on its presence.

Second: the agreement on "physician review solves it" may be softer than it appears. All four positions treat mandatory review as the governing safeguard. But the review assumption contains a hidden load: that physicians reviewing AI-drafted notes will catch errors at the same rate they would catch errors in notes they wrote themselves. There is behavioral evidence from adjacent domains --- aviation, radiology, financial audit --- that human review of machine-generated output is systematically less rigorous than independent human generation. If that pattern holds in clinical documentation, review is a weaker control than the panel assumed. None of the four tested this. It is the most consequential shared blind spot.

The strongest counter-argument --- that AI documentation use should be restricted or paused until longitudinal deskilling data exists --- does not prevail on current evidence. But it is not dismissible either. It holds as a legitimate minority position.

WHAT THIS MEANS

This question has one clear answer, but the answer carries a condition that is more load-bearing than the panel collectively acknowledged.

The consensus is correct: AI should be used as a drafting assistant, with physician review and sign-off on every note. The efficiency gains are real, the burnout crisis is real, and the tools are already deployed at scale. The question is not whether --- it is how.

The unexamined load-bearing condition is the review quality assumption. If physician review of AI-generated notes is treated as a checkbox rather than a genuine verification act, the safety architecture fails. The time saved by AI drafting must be partially reinvested in rigorous review --- particularly of the assessment, plan, medication reconciliation, clinically important negatives, and any finding that would change a treatment decision. A faster note that is wrong is worse than a slower note that is right.

The deskilling concern is real and open. It does not yet have the evidence base to override adoption. It has enough evidence to require monitoring.

BOTTOM LINE

Doctors should use AI to write clinical notes. The answer is yes.

The condition is not "with caution" in the vague sense. The condition is specific: the physician must review the AI draft as if it were written by an unsupervised junior trainee --- with the assumption that it contains at least one error --- not as a polished product awaiting signature.

The decisive factor is the review quality assumption. Every safety argument the panel made depends on physician review being genuine. If review degrades into rubber-stamping, the conditional yes becomes a no.

What would change this verdict: longitudinal data showing measurable decline in diagnostic reasoning among consistent AI-scribe users, or systematic evidence that AI draft review produces materially higher error pass-through rates than independent physician documentation.

The actionable directive: adopt AI scribes with institutional protocols that treat review as a clinical act, not an administrative step. Monitor error rates in AI-assisted notes against baseline. Build in mandatory training on where AI hallucination clusters in clinical documentation. Do not treat time savings as fully captured until review quality is measured.