Legislator Brief · one page · drafted to be printed and carried into a meeting

AI medical record integrity

A physician should be accountable for what she adopted, not for pretending she typed it.

Object type

Legislator Brief

Label

MODEL LEGISLATION

Status

OPEN FOR CRITIQUE

Jurisdiction

California / portable

Domain

Artificial intelligence & medicine

Baseline verified

2026-08-30

Issue

Generative systems draft clinical text that a physician then attests to as her own authorship, and nothing in the finished record distinguishes what she decided from what a model produced.

Why now

Ambient documentation is being deployed at system scale ahead of any provenance requirement, and the first contested cases will be decided on records that cannot answer the question.

Current law

Medical-record content and retention are governed by state record requirements and by electronic-record certification criteria; neither requires that machine-generated text be identifiable as such. FDA's clinical-decision-support policy addresses device status, not authorship.

Policy gap

Attestation law assumes a human author. No instrument requires provenance capture at generation, model-version recoverability, or a face-of-the-record distinction between physician and machine text.

KPSGILL recommendation

Require machine-readable provenance for AI-generated clinical text at the point of generation, and move physician responsibility from authorship to clinically material adoption.

Who can act

California Legislature (Health and Safety Code)ONC / electronic-record certificationMedical Board of California (guidance)

Cost

Analysis, not projection. Cost falls principally on record vendors as a one-time certification and engineering burden; provenance is written by software, so marginal clinician cost approaches zero. State cost is limited to rulemaking and certification oversight.

Trade-offs

Expected direction of effect if the recommendation is adopted as drafted. KPSGILL analysis.
DimensionDirectionBasis
Patient safety increaseTraceable text lets error be located rather than argued about.
Physician burden decreaseAttestation shifts to what was actually reviewed.
Litigation risk± mixedDiscovery becomes more precise, which cuts both ways for the clinician.
Innovation~ uncertainCertification cost is real; a single standard is cheaper than fifty.
Administrative complexity decreaseOne provenance field replaces institution-by-institution policy.
Privacy increaseAudit logs are separated from the clinical record under their own retention rule.

Who is affected

KPSGILL impact analysis. These are not claimed endorsements or stated positions of any organisation.
GroupExpected impactWhy
PatientsfavorableCan learn how their record was produced; no new consent burden.
Physiciansstrongly favorableRemoves attestation to authorship they do not hold; narrows exposure to adoption decisions actually made.
HospitalsmixedProcurement and audit obligations rise; liability allocation clarifies.
Technology vendorsunfavorableProvenance capture, version identity and retention become conditions of certification.
InsurersmixedCleaner evidentiary record in coverage disputes; no new duty.
GovernmentfavorableBoards and courts get an evidentiary basis they currently lack.

Policy options

Option A — status quo

Authorship fiction persists; boards and courts resolve the first cases on unrecoverable records.

Option B — limited reform

Institutional policy and voluntary vendor disclosure, with no machine-readable requirement. Cheap, unenforceable, and invisible to a later reader.

Option C — structural reform

Statutory provenance duty on holders plus a certification criterion, with retention and discovery rules attached.

Option D — KPSGILL preferred · preferred

Certification criterion carries the technical duty; statute carries the attestation relief and the retention rule. The vendor builds it once; the clinician stops attesting to fiction.

How we would know it worked

  • Share of certified record systems capturing generation-time provenance
  • Share of AI-drafted notes with recoverable model and version identity
  • Board cases in which provenance was available to the adjudicator
  • Clinician time spent on attestation language

The five-physician practice

A five-physician practice buys a certified product and inherits compliance. The duty is deliberately placed where the software is built, not where it is used.

Next decision point

The next electronic-record certification rulemaking cycle, and any state record-content bill introduced in the 2027 session.

Model language and sources

Model statutory or regulatory language, the documentary baseline it rests on, the strongest arguments against the proposal and the KPSGILL responses to them are on the full page: Model AI Medical Record Integrity Standard. Related briefs are indexed at Legislator Briefs.