When an AI Medical Scribe Becomes a Medical Device

A scribe that only transcribes is a documentation tool. A scribe that tells the clinician what the patient probably has is doing something else. Australia’s regulator has now found that most of the scribes it examined already cross that line. Here is how to regulate what these systems do, not what they are called.

Published August 9, 2026 · Rewritten 19 October 2026 · Record through October 19, 2026 · SOURCE-VERIFIED · EDITORIALLY REVISED · HUMAN REVIEW PENDING

The record this week

On October 16, Australia's Therapeutic Goods Administration published the outcome of its review of digital medical scribes.

The TGA scanned the market and found about one hundred suppliers. Forty answered a voluntary survey, and the TGA examined twelve products in detail. Of those twelve:

eight contained functions likely to meet Australia's legal definition of a medical device;
two needed more information before the TGA could say;
two appeared unlikely to qualify.

None of the twelve suppliers had identified its product as a medical device. Some relied on disclaimers stating that their products were not medical devices or were not intended for diagnosis or treatment. The TGA's answer was that such statements "are not sufficient" to decide the question.

The TGA was careful about what it had not done. It "did not test any digital scribe products, check the safety and performance, or assess compliance with detailed medical device regulatory requirements". It did not assess privacy or data protection, and it did not name the eight products. The finding is about function: what the software does with what it hears. It is not a finding that any product is unsafe or that any supplier has broken the law.

Three days later, on October 19, the comment period closed on the U.S. Food and Drug Administration's discussion paper on generative-AI-enabled medical devices (docket FDA-2026-N-7874). The paper is a request for feedback, not guidance. In mid-October the UK's MHRA published a 2030 strategy that, as reported by HTN Health Tech News, commits it to a modern regulatory framework for AI as a medical device. The strategy includes expanded predetermined change control plans and the possible use of master files for foundation models. It is a strategy, not law. Press accounts have also described Japanese draft review considerations for generative-AI devices; KPSGILL has not located that primary document and does not rely on it here.

None of these is a rule about AI scribes. Together they show where the line is being drawn.

Four things a "scribe" can do

The word scribe now covers four different activities, and each carries a different risk.

1. Transcription. Converting speech to text. The output is a record of what was said. Errors are errors of fidelity: a misheard drug name, a dropped negation ("no chest pain" becoming "chest pain"). These errors are serious, but the clinician can detect them by comparing the text with their own memory of the visit.

2. Documentation assistance. Reorganizing what was said into a note: history, examination, assessment, plan. The software now chooses what to include and where. Omission becomes a risk, because a symptom mentioned once may not survive summarization. The output still claims to reflect the encounter rather than add to it.

3. Clinical inference. Producing content the clinician did not say:

a suggested diagnosis or differential;
a flagged abnormal finding;
a "consider ordering" prompt;
a risk score;
a medication suggestion;
an assessment the clinician never voiced.

This is where Australia's line falls. The TGA treats software that generates clinical conclusions, recommendations or other outputs potentially affecting patient care as potentially a medical device. Software limited to transcription or administrative documentation generally is not.

4. Downstream action. Drafting orders, populating problem lists, proposing billing codes, or prioritizing which patients need follow-up. The output stops being text a clinician reads. It becomes structured data that other systems act on: order queues, claims, registries, quality dashboards.

Products can move between these categories as features are added; the TGA noted that existing products launched new functions during 2026. A feature added in a quarterly update can carry a product from category 2 into category 3. Neither the clinician nor the purchasing health system may notice. That is the regulatory problem in one sentence.

Why the marketing label cannot be the test

A vendor has every incentive to describe its product as administrative. Administrative software avoids premarket review, device-registration fees, postmarket reporting duties and, in many systems, the liability that attaches to a medical device. Calling a product administrative costs the vendor nothing; accepting device status costs a great deal.

The TGA's refusal to accept disclaimers is therefore the most consequential sentence in its report. If the label decided the question, every product would carry the cheapest label. The review assessed functionality alongside intended purpose, documentation and marketing, rather than relying on the supplier's label. It applies that approach to a category whose functions change faster than its marketing.

The U.S. framework reaches a similar conclusion through a different route. The Food, Drug, and Cosmetic Act excludes certain clinical decision support software from the device definition, but only if, among other conditions, the clinician can independently review the basis for a recommendation (FD&C Act §520(o)(1)(E), 21 U.S.C. §360j(o)(1)(E)). Analysis: a generative system that produces an assessment without showing what evidence it weighed may struggle to meet that condition. FDA's discussion paper does not resolve how these conditions apply to generative output. Its proposed "competency-based" approach (benchmarking plus clinical confirmation) and risk-proportionate postmarket monitoring are offered for comment, not adopted.

What goes wrong, function by function

Fidelity. Transcription errors propagate. A note signed in error becomes the source document for the next clinician, the insurer, the disability adjudicator and, sometimes, the courtroom.

Omission. Summarization is selection. A model optimized to produce concise notes will tend to drop information that appears once and seems peripheral. Those are often exactly the details (a passing mention of suicidal thoughts, a supplement, a fall) that matter most later.

Hidden inference. When a note includes an assessment the clinician did not voice, two questions arise that the record cannot answer: did the clinician adopt it, or merely fail to delete it? The attestation ("I have reviewed and agree") was written for human-drafted notes. It does not distinguish between the two.

Coding drift. Software that proposes billing codes from the note it wrote has a structural conflict. A system that both documents and codes can, without anyone intending it, document toward the code. That is a payment-integrity problem before it is a clinical one.

Version drift. The same product can behave differently after an update. If the record does not show which model version produced a note, no later reviewer can reconstruct why the note says what it says.

Provenance loss. Once a generated note is signed, the line between what the patient said, what the clinician observed and what the model inferred usually disappears from the record.

Privacy and vendor access. Ambient capture means continuous recording of the most sensitive conversations in medicine. The TGA expressly did not examine privacy. Its silence on that point is a gap, not a clearance.

What the clinical evidence shows

The best available studies, cited below, are encouraging on burden and silent on safety. A 2025 multicenter pre/post quality-improvement study of 263 ambulatory clinicians across six health systems reported lower self-reported burnout after ambient-scribe adoption. A 2025 qualitative study of 22 physicians found generally positive experiences with workload and patient engagement, alongside concerns about accuracy and editing.

Neither study measured diagnostic error, omission rates against a reference standard, or downstream harm. "Time in note" fell. Whether the note is more or less reliable as a clinical and legal document is not known from these data.

Who bears the risk now

The clinician signs the note and, under current professional standards, owns its content. A medical board will ask what the physician knew and attested. It will not ask what the vendor's model produced.
The health system chose the product, configured it and often mandated it. Analysis: procurement contracts can leave the signing clinician carrying responsibility for output while limiting the vendor's own liability.
The vendor controls the model, the update schedule, the default settings and the logs. It usually has the least legal exposure.

That allocation runs opposite to control, and it is the strongest argument for regulating function rather than label.

KPSGILL position

Regulation should follow what an AI documentation system does, what clinical influence it exerts, and who controls the resulting risk — not the vendor's description of the product.

In practice, KPSGILL proposes a California standard with five working parts. It is set out as model text in the Reform revision of the Model AI Medical Record Integrity Standard (see the Reform revision).

Functional classification. Each deployed feature is classified as transcription, documentation assistance, clinical inference or downstream action. Classification is by function, documented by the deploying health system, and revisited at every material update.
Visible provenance. Generated notes mark machine-inferred content (assessments, differentials, flags, codes) separately from transcribed content. The signed record carries the model name, version and generation date.
Clinician control. The clinician can inspect and edit all generated text before signature. No material content is altered after signature without a logged amendment. A clinician who overrides or deletes machine output faces no productivity penalty for doing so.
Change notification and incident reporting. Vendors notify deploying organizations of material model changes before deployment. Organizations maintain an incident log of documentation errors that reached a signed record and report serious incidents to an identified state authority.
Procurement accountability. A health system that mandates a documentation tool shares responsibility for its foreseeable failure modes. Contract terms cannot shift that responsibility entirely to the signing clinician.

The strongest objections

"This will slow adoption of tools that reduce burnout." Partly true. The answer is proportionality. Pure transcription carries obligations close to zero. The requirements attach only where a product generates clinical inference or downstream action, which is where the TGA found most of the products it examined (8 of 12) already sit.

"Clinicians already review and sign; that is the safeguard." Review is a safeguard only if the reviewer can see what needs review. Attestation of an undifferentiated note does not distinguish verification from failure to notice.

"Federal law may preempt state rules for devices." For products that FDA regulates as devices, state requirements that differ from or add to federal device requirements face preemption risk. The proposed standard is therefore written as a regulation of deploying health systems and professional practice: record integrity, provenance and procurement. It is not a design requirement for devices. That distinction must be maintained in any drafting.

"Small practices cannot run an incident-reporting program." The obligations should scale. A solo practice using a vendor product needs a log and a way to report. Classification and change-management duties rest mainly on the vendor and the larger deploying organization.

What to watch

TGA compliance actions and its promised guidance.
Whether FDA's next CDRH document on generative-AI risk and evidence (listed as "under construction" for FY2027) addresses documentation tools.
MHRA's implementation of expanded change-control plans.
Whether California's Medical Board treats generated documentation as a disciplinary question. See Forecast F-02 and the existing forecast that board cases will precede board guidance.

Sources and Authorities

Each source below was verified against the official publisher, current through August 9, 2026. Laws, proposed rules, and agency pages change; every link is re-opened live at deployment, and time-sensitive requirements should be checked against the current official source.

PubMed — Randomized Trial of Ambient AI Scribes

JAMA Network Open — Multicenter Ambient AI Scribe Quality-Improvement Study

JAMA Network Open — Physician Perspectives on Ambient AI Scribes

AHRQ Digital Healthcare Research — Digital Scribes

HHS OCR — Software Vendors and Business Associate Status

Sources added in the October 19, 2026 rewrite. The TGA page refused automated access; its content was verified through The Medical Republic (October 16, 2026), and Pulse+IT’s paywalled report confirms the 8-of-12 headline. The MHRA strategy is cited as reported by HTN Health Tech News; the primary MHRA document has not yet been read.

TGA digital scribe review outcomes (Primary (not retrievable), 2026-10-16; checked 2026-10-19)

The Medical Republic report on TGA review (Secondary (quotes report), 2026-10-16; checked 2026-10-19)

Pulse+IT report (Secondary, 2026-10-16; checked 2026-10-19)

FDA GenAI device discussion paper page (Primary, 2026-08-18; checked 2026-10-19)

Regulations.gov docket FDA-2026-N-7874 (Primary; checked 2026-10-19)

FDA GenAI discussion paper PDF (Competency-Based Approach, sec. V) (Primary, 2026-08-18; checked 2026-10-20)

CDRH FY2027 proposed guidances (Primary, 2026-10-01; checked 2026-10-19)

HTN: MHRA 2030 strategy (Secondary, 2026-10-15; checked 2026-10-19)