AI scribe drafts can read more certain than the visit was, a 2026 study finds

A 2026 preprint posted to arXiv put a number on something coders have half-suspected about AI-drafted notes. When clinicians rewrote the drafts an ambient-AI scribe produced, they more often added words that soften a statement than removed them. Researchers at the University of California, Irvine compared each AI draft against the clinician's signed note across 62,811 paired note sections, drawn from two commercial ambient-AI platforms piloted in outpatient clinics. On balance the edits leaned toward more uncertainty, which the authors read as clinicians pulling back drafts they judged too definitive. The plain reading for a coder: the draft tends to sound surer than the clinician was.

What the study measured

The team compared the AI draft and the signed note section by section across up to four parts of a note (history, exam, assessment and plan, and results), then counted how a curated list of 536 hedging terms moved between draft and final text. The analysis covered two vendors, anonymized as Vendor A and Vendor B, and both primary care and specialty encounters. This is a preprint, so it has not yet completed peer review, and every figure below is the authors' reported result.

Which way the edits went

The study's net measure of edit direction points toward uncertainty. Among the sections where a clinician made a hedging-related directional edit, more shifted the wording toward less certainty than toward more, and the authors report that net tendency as statistically significant. It applied to a minority of edits, though. Most note sections carried no hedging-related change at all, so this is a signal in the margins, not a rewrite of every note.

The single most common word substitution the study logged was "likely" softened to "possibly." The paper catalogs more than a thousand such substitutions, in both directions and sorted by frequency, so read that list as the texture of the editing rather than proof that every change ran one way. The direction comes from the net-score analysis, not the word list.

The authors' own framing is the useful part for a coder. They write that tentative phrasing should reflect whether a diagnosis is confirmed or still under consideration, and whether supporting evidence from labs, imaging, or pathology is present. That's a coding standard stated in clinical language. They describe the pattern as likely a cross-vendor issue rather than one platform's quirk, since both platforms they studied trended the same way.

Why an over-certain draft is a coding problem

Read the finding against the FY2026 ICD-10-CM Official Guidelines. The rule for the professional and physician-office record is explicit: you do not code a diagnosis documented as probable, suspected, questionable, rule out, compatible with, consistent with, or working diagnosis. You code the sign, the symptom, or the reason for the visit instead. The inpatient rule is the opposite, and that split is what makes an over-certain draft a live risk in the outpatient and physician-office setting where much risk-adjustment coding happens.

So the certainty word in the note is load-bearing. "Possible sepsis" in a clinic note does not support a sepsis code; you fall back to the documented signs. If an AI draft renders that same clinical picture as "sepsis," flat, and it is signed without the qualifier the clinician would have added, the note now reads like a confirmed diagnosis the encounter never established.

Where risk adjustment raises the stakes

Much of risk-adjustment coding happens in the outpatient and physician-office record, which follows the outpatient rule. A condition that maps to an HCC has to be supported by the record, and RADV review exists to confirm exactly that. A diagnosis that reads as confirmed but was never clinically established is not supported documentation, and unsupported HCCs are what a RADV audit recovers.

It's the mirror image of the scribe risk coders already watch, where the draft drops a chronic diagnosis and costs a recapture. One failure under-documents the chart. This one over-states it, and both land on the coder to catch.

That's the gap between how the study frames its finding (a clinician editing habit) and why it lands on a coder's desk. The clinician who catches the over-certain draft and softens it has closed the gap. The coder is the backstop for the note where that edit didn't happen.

What the study did not show

Be precise about what this is and isn't. The study measured notes clinicians did revise, and the shift toward uncertainty is evidence that clinicians treated the drafts as too definitive. It didn't measure coding outcomes, it didn't count how many over-certain drafts were signed without correction, and it didn't break the finding out by the assessment section specifically, where a diagnosis actually lives. The coder-facing risk (an uncorrected over-certain note reaching the claim) is a reasonable inference from the pattern, not a number the study reports. Treat it as a reason to check, not a measured error rate. The work also carries the usual preprint caveats: a single health system, no patient or clinician characteristics, and no adjustment for which vendor was used in which specialty.

Where the coder still owns the code

The study's own recommendation (make tentative phrasing match whether the diagnosis is actually established) is a coder's reflex, not a software feature. When you code from an ambient-AI or scribe-generated outpatient note, the certainty of a diagnosis is a claim to verify against the rest of the record, not a fact to inherit. Does the assessment's confidence match the labs, the imaging, the exam? Is there MEAT behind the condition, or just a confident sentence? When the certainty outruns the evidence, that is a provider query, the same call a coder has always made, now aimed at a draft a machine wrote first. That judgment is the part no ambient scribe in this study could do on its own.