Signing AI-Scribed Notes: What Your Signature Attests

An ambient AI scribe listens to the session, drafts the note, and presents it for signature. The review step is where the time savings are either kept or lost, and it is also where legal responsibility settles. Under Medicare's rules, the clinician who signs an AI-scribed note affirms that the note adequately documents the care provided, whether a person or a model drafted it. This post sets out what CMS actually requires, the specific ways AI-scribed psychiatry notes go wrong, the early evidence on reducing those errors, and a short checklist to run before signing.

What CMS requires when a scribe drafts the note

CMS addressed AI scribes directly in its fact sheet on Medicare signature requirements, MLN905364 (July 2025): "If you use a scribe, including artificial intelligence technology, sign the entry to authenticate the documents and the care you provided or ordered." The fact sheet adds that you do not need to document who or what transcribed the entry, and the scribe is not required to sign ^[1].

The Medicare Program Integrity Manual, chapter 3, section 3.3.2.4, states what the signature itself means: the treating practitioner's signature on a note "indicates that the physician/NPP affirms the note adequately documents the care provided," and the manual applies the same practitioner-concurrence requirement when AI technology is used to capture the transcription of medical record entries ^[2]. CMS did not create a separate, lighter standard for machine-drafted notes. The signer owns the content of the note, line by line, in the same way as a note they typed themselves. Related compliance groundwork, including what a BAA covers and does not, is in our HIPAA and BAA guide.

Where AI-scribed psychiatry notes go wrong

The failure modes are consistent across the commentary and early audits, and several are specific to psychiatric practice:

• A differential charted as a diagnosis. A possibility raised aloud during assessment ("this could be major depression, could be adjustment") is written into the assessment as a settled diagnosis. A caregiver advocate writing in KevinMD described this as turning "a maybe into a diagnosis," and proposed a psychiatry-specific review step before signing for exactly this reason ^[3].
• Collateral merged into the patient's history. Statements from a spouse, parent, or other collateral source are folded into the patient's own reported history, losing who actually said what ^[3].
• Questions blurred with answers. Something the clinician asked is charted as something the patient reported, or a topic raised is recorded as a finding elicited.
• Invented findings. The note contains examination elements, observations, or history that were not elicited in the session, written in the same confident register as the accurate content ^[4].
• Copy-identical notes. Successive session notes repeat the same mental status examination, risk language, and plan with little change, even when the sessions differed ^[4].
• Complexity drift. Documented complexity creeps upward over time relative to the care actually delivered, which matters because the signed note supports the billed service ^[4].
• Batch signing. Notes are signed in a stack at the end of the day or week, when recall of any individual session is weakest and review is thinnest ^[4].

Psychiatry is exposed on each of these because so much of the record is reported speech, attributed history, and clinical inference rather than measured values. A merged collateral statement or an upgraded differential is not a typo; it changes the diagnostic record that follows the patient.

What the evidence shows about reducing these errors

One recent evaluation measured two of these failure types directly. In BMJ Innovations (October 2026), researchers evaluated the TORTUS ambient scribe within the MHRA's AI Airlock regulatory sandbox across 5,719 consultations. With both of the system's safety mechanisms switched on (one designed to strip out content with no support in the consultation, the other to catch drift beyond what was discussed), the share of note content that was unevidenced fell from 3.3% to 1.6%, and measured drift risk fell from 2.9% to 1.2% ^[5].

Three caveats apply before generalizing those numbers to psychiatry. The cohort was general care, not psychiatry, so applying the rates to psychiatric encounters is an extrapolation. Three of the four authors are at TORTUS AI, the vendor whose product was evaluated. Drift was scored by an LLM acting as judge, not by independent clinician review. The defensible reading is narrower than the headline: layered safety checks reduced measured error rates in one large sandbox evaluation, and the residual error rate was not zero, which is why the signature review remains the clinician's step, not the vendor's.

A pre-sign checklist for behavioral-health clinicians

None of these checks requires re-reading every line. They target the places where AI-scribed psychiatry notes most often diverge from the session:

• Diagnoses: every diagnosis in the assessment is one you actually assessed and intend to stand behind; differentials are labeled as differentials, not carried as confirmed.
• Whose history: collateral information is attributed to its source (partner, parent, prior record) and has not been merged into the patient's own reported history.
• Reported speech: statements in quotation or paraphrase are things the patient actually said; your questions have not been charted as their answers.
• Medications: every start, stop, and dose change in the plan matches a decision actually made in the session, including controlled-substance details.
• Risk statements: suicidal and homicidal ideation, intent, plan, and protective factors reflect this session's assessment, not language carried forward from a previous note.
• Findings: mental status elements and observations in the note were observed or elicited in this encounter, not generated to complete a template.
• Signing pattern: notes are reviewed and signed per encounter, close to the session, not approved in an end-of-week batch.

Bring sign-off and demo questions to vendor conversations with our buyer's checklist.

The bottom line: CMS treats your signature on an AI-scribed note as an affirmation that the note documents the care you provided, so the pre-sign review (differentials, attribution, medications, risk) is not clerical cleanup; it is the clinical act the signature certifies.

Sources

1. Complying with Medicare Signature Requirements (MLN905364). Centers for Medicare & Medicaid Services. https://www.cms.gov/mln905364-complying-with-medicare-signature-requirements (opened 2026-10-26)
2. Medicare Program Integrity Manual, Chapter 3, §3.3.2.4, Signature Requirements. Centers for Medicare & Medicaid Services. https://www.cms.gov/regulations-and-guidance/guidance/manuals/downloads/pim83c03.pdf (opened 2026-10-26)
3. Gillen, N. D. "AI scribes in psychiatry can turn a maybe into a diagnosis." KevinMD. https://kevinmd.com/2026/10/ai-scribes-in-psychiatry-can-turn-a-maybe-into-a-diagnosis.html (opened 2026-10-26)
4. Hillman, A. J. "The AI Scribe Wrote the Note. Your Signature Owns It." Medium. https://andrewjhillman.medium.com/the-ai-scribe-wrote-the-note-your-signature-owns-it-058328b5d2b1 (opened 2026-10-26)
5. Evaluation of the TORTUS ambient AI scribe in the MHRA AI Airlock regulatory sandbox (article bmjinnov-2026-001648). BMJ Innovations. https://innovations.bmj.com/content/early/2026/10/08/bmjinnov-2026-001648 (opened 2026-10-26)