The Scribe Was Graded. The Chat Box Was Not.

This describes the public record as of September 12, 2026. The report’s recommendations carry a target completion date of April 2027, so the position described here is a snapshot of an open corrective action, not a settled state.

On June 11, 2026 the VA Office of Inspector General published Report 26-00182-140, a review of generative AI chat tools in clinical use. The finding is a single sentence with a long shadow: VA identified Ambient AI Scribe as rights- and safety-impacting in 2024 and high-impact in 2025, and did not identify VA GPT or Microsoft 365 Copilot Chat as high-impact use cases.

Both sets of tools were being used to produce clinical documentation. One got pre-deployment testing, an impact assessment, ongoing monitoring and human oversight. The other got a policy analogy.

What the analogy was

VA leaders described the general-purpose chat tools as analogous to search engines and emphasized user-level responsibility. It is a reasonable-sounding position. A search engine returns links, a clinician decides what to do with them, and no one writes an impact assessment for a search box.

The Inspector General called the analogy flawed, and gave the reason in plain terms: unlike a search engine that finds links to websites, generative AI tools synthesize and transform sources to produce novel content. A search engine hands you someone else’s sentence. These tools write a new one, and in this case the new one went into a medical record.

The report’s own numbers close the argument. Of 135 prompts shared through VHA’s prompt library, 79 were clinical: 56 for clinical notes, 17 for summarization, 6 for something else. Drafting a clinical note from a clinical encounter is the job Ambient AI Scribe was designated high-impact for doing. Staff were doing it in the tool that was not designated, by typing patient information into a prompt.

Engagement was not marginal. The Chief AI Officer’s Teams site recorded 10,997 active users over 90 days and the National Artificial Intelligence Institute’s recorded 4,835.

The structural problem, which is not VA’s alone

It would be easy to read this as an agency that misjudged a call. That reading misses what actually happened, and the more useful version is uncomfortable for every federal program running general-purpose AI.

VA did grade a deployment. Ambient AI Scribe has a defined purpose, a defined input, a defined output and a defined place in the record, so it presents as a use case and M-25-21’s machinery has something to attach to. The high-impact determination, the impact assessment, the named risk acceptor: all of it needs a use case to hang on.

A general-purpose chat tool has no single use. It has tens of thousands of them, one per prompt, invented by users at the moment of typing. There is no use case to designate, so nothing gets designated, and the tool is classified instead by what it resembles. It resembled a search engine, so it was governed like one.

This is the failure mode that follows from getting the unit of governance right in principle and losing it in practice. We have argued before that federal policy has the better instrument, because it grades the deployment rather than the organization. This report is the sharper version of the same point. Grading the deployment only works if you can see the deployment. A product that is a thousand deployments wearing one name defeats the instrument, and defeats it quietly, because nothing shows up as unassessed. Nothing shows up at all.

Note what the determination turned on. Not stakes, not reach, not reversibility, but form. The question a program should be able to answer is the one M-25-21 actually asks, which is whether the output is a principal basis for a decision with significant effect on human health and safety. The question that got answered was what the interface looks like.

The part with no evidence trail

The report’s second finding is the one that should worry a program office more than the first.

There is no AI-specific reporting mechanism or labeling process to retroactively identify AI-generated documentation. The Joint Patient Safety Reporting system takes roughly 180,000 submissions a year. A search of it returned zero mentions of AI.

Zero is not a safety record. Zero is the absence of an instrument. If a clinician drafted a note with a general-purpose tool and something in that note was wrong, there is no field that says so, no marker on the record, and no way to go back through the corpus and find the others like it. The tool’s output entered the authoritative record and became indistinguishable from everything else in it.

That is the condition CISA’s Logging Reference Architecture legislates against for security operations, where section 10 requires named human review points, fallback procedures, and chain-of-custody separating event records from model-generated output, and states that AI output must not become the authoritative record. The same principle, applied to a clinical note, would have required a label. It is also what Law 5 is for. A right to intervene that cannot locate what to intervene in is a right on paper. You cannot stop, correct or recall what you cannot find.

And it bears on Law 1. The search-engine framing put responsibility on the user, which is a coherent position only while the tool is a tool. The moment it produces content that stands in the record under a clinician’s name, the question of whose output it is has an answer, and that answer needs an identity attached to it.

What the Inspector General asked for

Three recommendations, all to the Under Secretary for Health: review current use of generative AI chat tools and define permissible clinical uses, oversight responsibilities and risk mitigation with a plan for implementation; evaluate whether safeguards applied to high-impact tools such as Ambient AI Scribe should be adapted for chat tools used in clinical care and documentation; and integrate AI-related risk monitoring into existing patient safety programs, with staff trained to identify and report AI-related safety events.

The Under Secretary responded on May 19, 2026, concurring in principle with the first and concurring with the other two, with a target completion date of April 2027. VHA’s AI in Healthcare Workgroup is to develop guidance covering pre-deployment performance assessments, impact assessments, ongoing monitoring, use-case-specific education and transparency mechanisms, and the National Center for Patient Safety is to update patient safety reporting to track AI-specific risks and incidents.

Concurred in principle is worth noticing on the first one. It is the recommendation that would require VHA to decide what a general-purpose tool may and may not be used for, which is the hardest of the three and the only one that changes what happens tomorrow.

Three questions worth asking about your own program

The report is about VA. The structure is about everyone.

Which of your AI tools has no use case attached to it? Run the list. Anything that is a capability rather than an application has probably never been assessed, because there was no single thing to assess. That is the population at risk, and it is usually the most widely deployed.

If an output from one of those tools turned out to be wrong six months ago, could you find the others? If the answer involves interviewing people, you do not have an evidence trail. Labeling at the point of generation is cheap. Reconstructing it afterward is not possible.

Does your incident reporting have a field for this? A safety or quality system that has never received an AI-related report is more likely missing the field than missing the events.

None of this requires waiting for new guidance. M-25-21 already asks the question by effect rather than by interface. The work is applying it to the tools that did not arrive looking like a use case.