Healthcare AI Accountability · Physician essay and governance proposal

When the AI Scribe Becomes the Witness

Why peer review must examine how the medical record was made

  • Kanwar Partap Singh Gill, MD
  • Family medicine physician & independent policy analyst

Published

Public sources checked · Physician essay and governance proposal by the author · 14 min read

A signed clinical note on a review desk, with earlier versions and an audio waveform illustrating how the record was created.
Conceptual illustration. No real patient, clinician, product screen or record is depicted.

Imagine a quality-review meeting.

A chart appears on the screen. The assessment sounds confident. The plan says a specialist referral was placed. The physician's signature sits beneath it.

The reviewer cannot find a matching signed order in the records provided. A familiar question follows: Why did the physician document an action that apparently did not occur?

That is a legitimate question. There is another question the review should investigate alongside it: How did that sentence enter the signed record?

Perhaps the physician wrote it. Perhaps an ambient system changed a discussion about a possible referral into a statement that the referral had been placed. Perhaps the physician corrected an earlier draft, then another recording generated a new version. Perhaps the order exists in a system the reviewer has not examined. Perhaps the record accurately exposes a clinical or documentation failure.

These possibilities lead to different findings and different remedies. A credible review needs evidence that can distinguish them.

An AI-assisted note has acquired a second life. It helps describe the encounter, then becomes evidence used to evaluate the person who signed it.

Before the chart becomes a witness about the clinician, the review should establish what can be known about the chart.

The second life of a clinical note

The word "scribe" suggests someone quietly taking notes. Modern ambient documentation can involve several transformations between the conversation and the final record.

Microsoft describes Dragon Copilot and DAX Copilot generating diagnoses for assessment-and-plan content from ambient recordings and comparing them with diagnoses retrieved from Epic. Depending on alignment, generated diagnostic content can remain in the note, while retrieved diagnoses may appear without an associated plan. That describes a documentation and reconciliation process; it does not establish that the software independently diagnosed a particular patient. [1]

Microsoft also identifies omissions, hallucinated details, substituted clinical terms, and contradictions among possible limitations. Its transparency guidance requires clinicians to review and validate generated content before it enters the medical record. [2]

The physician's responsibility to verify the final record remains central. The additional question is what an organization can establish when verification appears to have failed.

A final note can support a finding about what was authenticated. It does not, by itself, reveal every step through which the disputed sentence passed. When those steps matter, a reviewer should examine them.

This matters to patients as much as clinicians. An investigation that attributes every discrepancy to the signer may miss a recurring software or integration problem. An investigation that attributes every discrepancy to AI may excuse an unsupported clinical decision. Both can leave the next patient exposed to the same failure.

The purpose of tracing authorship is to identify the failure accurately enough to correct it.

Related on this site: When the AI Scribe Starts Shaping the Medical Record

The correction that did not survive regeneration

Consider a second hypothetical case.

A clinician notices that a draft has turned an uncertain diagnosis into a definite one. The clinician corrects it. Later, an additional recording is added. A new version appears. The disputed wording returns, and the clinician fails to catch it before signing.

A reviewer who sees only the final note may conclude that the clinician never recognized the distinction. A version history could show that the distinction was recognized and that the correction did not survive regeneration. The final error still requires scrutiny, but the explanation and prevention plan change.

There is a concrete reason to test this sequence. Microsoft's note-version instructions state that additional ambient recordings create a new version from the combined transcripts and do not incorporate earlier manual edits or style changes. The same instructions describe version comparison and restoration. This documented behavior does not prove that any particular disputed note was affected. [3]

For the hypothetical clinician, the review would need to ask whether the final version was checked and whether the available workflow made regeneration and its consequences clear. The institution would need to examine training, interface behavior, and the transfer into the electronic health record.

"Did you review the note?" is therefore an incomplete operational question. A more useful question is: "Which version did you review, and what changed afterward?"

The relevant sequence includes the encounter, draft generation, correction, regeneration, transfer, authentication, and any later amendment. Different timestamps answer different questions.

That chronology can support accountability. It can also prevent a reviewer from attributing the wrong reasoning to a clinician.

Related on this site: The Medical Record Has More Than One Clock

When a checklist rewards a fictional encounter

There is another failure mode worth testing: documentation that improves its score by moving away from what happened.

Imagine a review checklist asking whether counseling, precautions, medication review, and follow-up appear in the note. An AI instruction tells the system to produce a complete note satisfying those criteria. The source material contains no account of counseling, yet the generated note says the patient was counseled and understood.

The words satisfy the checklist. The encounter has acquired an event that was never established.

This is why a documentation rubric should operate as a set of questions for verification. It should never become permission to fill factual gaps with the care that would have been desirable.

The distinction can be subtle. A model may identify a sensible test that the clinician never considered. Writing that the test "was indicated" can attribute a clinical judgment the clinician did not make. Writing that it "was recommended" adds a communication. Writing that it "was ordered" adds an action in the health record.

Each sentence makes a different factual claim. A clinically sensible suggestion does not acquire a past merely because software puts it in the past tense.

New suggestions can be useful in an authorized clinician-facing review workflow. Their status should remain visible until the clinician evaluates them. If a later review leads to a new recommendation, order, or patient contact, the record should preserve that later chronology.

The same principle applies to consent, chaperones, medication administration, and consultation. Generated prose cannot establish that an event occurred simply by describing it.

There is a further risk if an institution uses another AI system to evaluate the completed note. Without independent sources, that second system may reward the first system's unsupported additions as evidence of thorough care. This is a design risk to test, not a claim that a particular organization has adopted such a process.

Related on this site: UNFAITHFUL

Four questions hidden inside one quality score

A single question such as "Was the assessment and plan appropriate?" can conceal several distinct inquiries. I propose separating four dimensions whenever a material discrepancy makes the distinction necessary.

DimensionQuestion the review should answer
Encounter fidelityDoes the record accurately represent the history, findings, decisions, and uncertainty established for the encounter?
Clinical appropriatenessWere the actual clinical judgments and management reasonable in the circumstances?
Action completionWere the documented orders, communications, procedures, and follow-up completed to the stage claimed?
Provenance and verificationWhat is known about how consequential content was generated, changed, reviewed, and authenticated?

A note can faithfully document a poor clinical decision. A note can also describe an excellent plan that no clinician selected or carried out. A sound quality process needs to recognize both.

Return to the hypothetical missing referral. If the authenticated note says the referral was placed and the relevant complete order record establishes that it was not, the documentation discrepancy can be scored. The reviewer can separately evaluate whether referral was clinically appropriate and whether a care failure occurred.

If only an incomplete export was available, the reviewer should record that limitation before concluding that no order existed. A source limitation and a proven failure are different findings.

This is where a fourth review outcome becomes useful alongside "Meets," "Does not meet," and "Not applicable": "Unable to assess."

Apply "Unable to assess" to the unresolved question while recording any discrepancy already established. Identify the missing evidence, who will seek it, and when the finding will be revisited. The category must not erase an error in the signed record or quietly remove uncertain cases from a performance denominator.

The defensible finding may be: "A documentation contradiction was confirmed; whether the action occurred elsewhere remains unresolved; the source of the wording could not be established."

That is more informative than a single unexplained failure mark. It identifies what requires correction and what requires further investigation.

What an edit or rejection can actually prove

A clinician who changes more AI text may be correcting more errors. A clinician who changes very little may have received an excellent draft. Neither edit volume, by itself, establishes the quality of the review.

The same caution applies to accepting and rejecting suggestions. Microsoft's diagnosis-specificity instructions permit accepting, rejecting, or taking no action on a suggestion. The behavior of another feature or local integration must be examined separately. [4]

A rejected refinement might concern unsupported severity, a duplicate, or the wrong context. It should not automatically be interpreted as a considered decision that the patient did not have the broader condition. Equally, a clinically important warning still deserves assessment on its merits.

Microsoft's audit documentation lists Dragon Copilot Web events for note generation, note updates, and suggested edits presented, accepted, or rejected. Those event names identify possible evidence sources. They do not establish that every deployment retains the full text, that a particular event concerns a diagnosis, or that a log captures the clinician's reasoning. [5]

Before turning an event into an adverse conclusion, the reviewer should know what the event means in that system. Absence from an unvalidated log is a weak basis for claiming that an action never occurred.

The more useful measures are clinically material errors, whether they were corrected before authentication, what reached patient care, and whether confirmed problems were resolved.

When the review calendar outruns the evidence

The institution may retain the signed note long after some material used to create it is no longer readily accessible.

Microsoft's public instructions distinguish a 30-day window for accessing individual sessions in the application from a FAQ describing retention of audio, recognized text, transcripts, and flowsheet values for up to 90 days. [6, 7]

These statements do not establish that all evidence disappears on day 30, or that every recording can be retrieved until day 90. Application access, retention, export, and preservation arrangements require specific answers.

This matters when the review cycle is slower than the source-data lifecycle. For Health Center Program participants, HRSA's compliance manual describes at least quarterly QI/QA assessments using systematically collected patient-record information. It also leaves methodology and report design to the health center. It does not require waiting until quarter-end to investigate a material concern. [8]

A practical approach would use timely review of significant discrepancies and rolling samples, with quarterly reporting where applicable. The organization should know which artifacts remain available, who can retrieve them, and how an appropriate preservation request is handled.

Permanent retention of every clinical recording is not a prerequisite for useful review. Privacy, security, applicable duties, and the purpose of preservation all matter. Relevant material should remain within authorized processes, with an accountable owner for access and preservation decisions.

The absence of a recording also does not end the inquiry. Orders, results, messages, contemporaneous observations, and the clinician's account may still answer the question. Audio can be incomplete, and it cannot establish that an order was actually transmitted or a medication administered.

The review should explain what the available evidence can support and where its limits remain.

A second signature needs a defined meaning

The problem becomes more complicated when another clinician later reviews or countersigns the chart.

Did that person examine the patient? Discuss the case at the time? Review the final note afterward? Check selected orders? Conduct a periodic quality assessment?

Those activities provide different kinds of assurance. An attestation should describe the activity actually performed and any resulting action. A later signature cannot retroactively supply an earlier conversation that did not occur.

California also requires precision about professional roles. The Physician Assistant Board explains that routine physician review or countersignature of every PA chart is not generally required unless the practice agreement requires it, with specified disciplinary exceptions. The Board of Registered Nursing distinguishes traditional practice under standardized procedures from qualifying 103 and 104 NP practice. A generic "PA/NP supervision" rule can obscure those differences. [9, 10]

AI documentation safeguards should apply across professional categories. Additional consultation, supervision, and review should follow the person's actual authority, applicable agreement or procedures, privileges, and clinical circumstances.

Where a required consultation is material to care, the reviewer should establish what communication occurred. A generated sentence saying "discussed with the supervising physician" is itself a claim requiring support.

A review process that can explain its conclusions

There is empirical support for examining drafts and corrections separately. A 2026 prospective pilot report described 31 physicians generating 7,545 notes during a two-month 2024 pilot. Physicians evaluated 356 drafts; 19 of those evaluated notes, or 5.3%, contained errors they rated as posing serious or imminent risk if uncorrected. Most evaluated drafts were free of significant errors. These were selected, physician-rated drafts from one institution; the result was not a measured rate of patient injury or a current error estimate for every ambient product. [11]

For governance, the important question is what happens between a problematic draft and the record used for care. A detected and corrected draft error, an error remaining in an authenticated note, and an error that affects treatment call for different findings.

The following practices would make reviews more informative without requiring every encounter to become a forensic examination.

Identify the exact concern and record. Name the consequential statement, the authenticated version, and the relevant time. State whether the concern is about clinical judgment, factual documentation, action completion, or another defined requirement. Do not infer AI involvement from writing style.

Build a proportionate evidence set. Start with the final note and the records that can answer the specific question. Add drafts, version history, transcripts, recordings, or relevant logs when available and necessary. A routine review may need little beyond ordinary clinical records; a disputed transformation may require more.

Apply the applicable criteria consistently. Identify the policy, clinical expectation, and evidence supporting the finding. Account for encounter modality and scope. Evaluate historical work against the requirements applicable at the time. Avoid treating a difference between narrative and coded fields as automatically wrong; historical conditions and differentials can legitimately have different statuses.

Hear the clinician's explanation and test it. A response can identify missing records or a legitimate clinical rationale. It can also reveal an unsupported decision or a verification failure. Neither an untested accusation nor an untested explanation should settle a contested factual issue. The procedural rules governing any formal proceeding remain specific to its setting.

Correct the patient record and the process. Assign an owner and urgency to each material problem. Determine whether an error also reached structured diagnoses, orders, instructions, or another destination. Preserve the chronology of corrections. Patient-care follow-up should proceed when indicated even if the origin of the error remains unresolved.

Examine institutional contributions. A repeated defect should prompt review of configuration, training, workload, integration, and vendor performance. Representative samples should describe routine performance; cases selected because of known concerns should be reported separately. Combining them without explanation can distort comparisons among clinicians.

These recommendations fit the SAFER High Priority Practices guidance, which addresses shared AI safety responsibilities, ongoing monitoring, inventories of deployed systems, transaction logs, and processes for reporting and analyzing safety concerns. That guidance supports a governance approach; it does not make this proposed review method a universally mandated legal procedure. [12]

The institution's strongest objection deserves a direct answer: the clinician signed the note and was responsible for reviewing it. Agreed. The proposed process examines whether that responsibility was fulfilled and what else contributed to the outcome. A signature identifies an accountable professional. It does not complete the investigation.

Related on this site: Review a Clinical AI Tool Before Purchase, After Changes, and When Something Goes Wrong

The patient has the largest stake

For the patient, a false sentence can matter long after the review meeting.

In another hypothetical case, "the patient requested time to consider treatment" becomes "the patient refused treatment." A later reader may attribute the delay to the patient. If a quality review accepts that sentence without examining a credible discrepancy, the error gains another institutional endorsement.

The same can happen when a tentative diagnosis becomes definite or a proposed action becomes a completed one. A review can reinforce the error or help uncover it. Its value depends on how carefully it tests consequential claims.

Useful ambient technology can assist clinicians with documentation. Useful peer review can identify unsafe care. The relationship between them requires a clear account of how the record was produced and what it can establish.

The goal is a trustworthy record and a review that can explain its findings. That means recognizing competent correction, identifying missed verification, investigating system defects, and responding to actual patient-care failures with the evidence each conclusion requires.

When the AI scribe becomes the witness, medicine should be prepared to examine how its testimony was written.

Public source register

All links were checked on 18 November 2026. Product pages describe their stated scope and may change. They do not verify a particular organization’s configuration.

  1. Microsoft Learn. Intelligent diagnoses retrieval from Epic. Described generation and reconciliation of assessment-and-plan diagnoses.
  2. Microsoft Learn. Transparency white paper. Intended role, clinician validation, and stated limitations.
  3. Microsoft Support. Work with note versions in Dragon Copilot. Regeneration, prior manual edits, version comparison, and restoration.
  4. Microsoft Support. Diagnosis specificity suggestions. Available dispositions for that feature.
  5. Microsoft Learn. Audit log activities, Dragon Copilot Web activities. Defined event types and their stated scope.
  6. Microsoft Support. Create and manage your sessions. The stated 30-day application-access window.
  7. Microsoft Learn. Frequently asked questions about Dragon Copilot. Retention of specified data for up to 90 days.
  8. HRSA. Health Center Program Compliance Manual, Chapter 10, Quality Improvement/Assurance. Quarterly assessments and methodological discretion.
  9. California Physician Assistant Board. SB 697 Frequently Asked Questions, questions 10 and 11. Chart-review and supervision distinctions.
  10. California Board of Registered Nursing. Assembly Bill 890. Traditional practice and the 103 and 104 categories.
  11. Taylor SL and colleagues. Quality of Clinical Notes Created by Ambient Listening Generative AI: Pragmatic Prospective Pilot Study. JMIR Medical Informatics. 2026;14:e86474. Published April 17, 2026; pilot conducted July-August 2024.
  12. ASTP/ONC. SAFER High Priority Practices, recommended practice 1.5, PDF pages 14-15. AI governance, monitoring, transaction logs, and safety reporting.

Editorial and informational notice: this analysis of health care, technology, consumer and legal policy is published for education and public awareness. It is not individualized medical, legal or financial advice and does not replace your own clinician, lawyer, bank or benefits office. Court orders, settlements and filings are described for what they decide; allegations are not findings. Kanwar Partap Singh Gill, MD is a physician, not a lawyer, and this site is not a law firm. In an emergency, call 911.

Physician essay and governance proposal by Kanwar Partap Singh Gill, MD. Public sources checked 18 November 2026; facts checked, fully reviewed. The scenarios are hypothetical; the review safeguards are recommendations, not existing legal requirements. Corrections: editorial standards.

Search KPSGill.com

Search the full publication — patient guides, policy, investigations, enforcement records and research.