“Like a Search Engine”: How the VA Let AI Chat Tools Into Clinical Care Without High-Impact Safeguards

Picture a busy clinic afternoon. A clinician has a visit transcript, a stack of results and fifteen minutes before the next patient. On the work laptop there is a chat tool that can turn all of it into a tidy draft note. It is quick, and it reads well. Nobody asks whether that tool has been tested for this job, and if the draft quietly gets something wrong, there is no box on any safety form to say that an AI was involved.

That scene is illustrative. The gaps in it are not. They are what the US Department of Veterans Affairs’ inspector general documented in June. Its report, Review of Generative Artificial Intelligence Chat Tools for Clinical Use (Report 26-00182-140, 11 June 2026), found that VA had given clinical staff general-purpose AI chat tools “without implementing oversight or safeguards” and without a way to track the errors they might cause (VA Office of Inspector General [VA OIG], 2026, p. 8).

Two chat tools, widely used

VA provides two general-purpose chat tools for work that may involve patient information: VA GPT, which VA built itself, and Microsoft 365 Copilot Chat (VA OIG, 2026, p. 1). Staff supply a prompt along with clinical information, and the tool produces text. Among the clinical prompts the inspector general reviewed, drafting clinical notes and summarizing patient care were the most common uses (p. i).

How widely are they used? VA does not know. The report says there is “no means to measure” how broadly the tools are used for clinical care (p. 4). What the inspectors could see was engagement: over 90 days ending 6 January 2026, a VA GPT feedback and discussion site on Microsoft Teams had 10,997 active users, and a second AI site had 4,835 (p. 4, Table 1). On a prompt-sharing app built by two staff members, the inspectors found 135 shared prompts. They classed 79 as clinical: 56 for writing clinical notes, 17 for summarization and 6 for other purposes (pp. 4–5). VA does not centrally curate or evaluate prompts like these (p. i).

As one VA AI leader told the inspectors, the tools “are not purpose-built as clinical solutions” (p. 5). The report points to a published study of AI-drafted clinical documentation that found a hallucination rate of 1.5 percent, with 44 percent of those errors considered major, meaning they could have affected diagnosis or treatment if left uncorrected (p. 5, citing Asgari et al., 2025).

“Analogous to search engines”

The heart of the report is a classification decision. Under the White House Office of Management and Budget’s April 2025 memorandum M-25-21, federal agencies must identify “high-impact” AI, meaning AI whose output is a principal basis for decisions with a significant effect on, among other things, human health and safety (Office of Management and Budget [OMB], 2025, p. 19). The memo presumes that AI used for “patient diagnosis, risk assessment, or treatment” is high-impact (OMB, 2025, p. 21). Once a use is high-impact, a set of minimum safeguards applies.

VA did not classify VA GPT or Copilot Chat that way. In interviews, VA leaders described the tools as “analogous to search engines” and emphasized the responsibility of individual users (VA OIG, 2026, p. ii). The inspector general disagreed plainly: “this analogy is flawed” (p. 6). A search engine points you to sources. A generative tool synthesizes and transforms them into new content, such as a medical record entry drafted from a visit transcript.

The contrast inside VA makes the point sharper. VA is piloting Ambient AI Scribe, a tool that listens to clinical visits and drafts notes. VA treats the scribe as high-impact and applies the corresponding safeguards (p. 5). Those include pre-deployment testing, an AI impact assessment, ongoing monitoring, training, additional human oversight, remedies or appeals, and feedback from users and the public (p. 6). The chat tools received none of this, even though, the report notes, most of the clinical prompts staff were sharing did something similar to what the scribe does (p. 6). Same job, two different rulebooks, decided by the label on the tool rather than the work it was doing.

Nowhere to report it

The second finding is quieter and, to me, more troubling. VA records medical errors, close calls and near misses in the Joint Patient Safety Reporting system, which it shares with the Defense Health Agency. According to VA’s patient safety AI lead, the system receives about 180,000 submissions a year (VA OIG, 2026, p. 7). The inspectors searched it for any mention of AI. They found none.

That is not evidence that nothing has gone wrong. It is evidence that the system had no way to see it. There was no AI category on the reporting form. In 2025, the National Center for Patient Safety asked the Defense Health Agency to add AI to the drop-down list of causes, but no formal decision was reached (p. 7). Nor was there any way to label or trace AI-generated documentation after the fact (p. iii). The AI teams and the patient safety center had barely coordinated (pp. 6–7). The patient safety lead’s summary was candid: “It’s very tricky. We don’t have the best answer yet” (p. 7).

Speed was the stated priority

Why did this happen? VA leaders described prioritizing rapid access to the tools over safeguards they considered impractical. One told the inspectors the guidance was to “minimize unnecessary levels of bureaucratic oversight” and push risk management to the lowest reasonable level (VA OIG, 2026, p. 8). The inspector general acknowledged that this matches the same OMB memo’s call to remove barriers to innovation, but noted that the memo also requires risk management “proportionate to the anticipated risks from its use” (p. 6; OMB, 2025). Removing red tape and classifying risk honestly are not opposites. The memo asks for both.

There is something worth crediting here, too. After the inspector general’s preliminary memorandum in January 2026, VA added patient safety representation to its AI Assessment Subcommittee and began working with the Defense Health Agency to enable AI-related event reporting (VA OIG, 2026, pp. iii, 7). The patient safety center plans an AI-specific reporting category and an education campaign to help staff recognize AI-related events (p. 7). The inspector general, for its part, disclosed that its own team used Copilot Chat with tested prompts to review inspection data, then checked the output against the sources (p. 10).

What VA agreed to

The inspector general made three recommendations to VA’s Under Secretary for Health: define permissible clinical uses, oversight and risk mitigation for the chat tools; evaluate whether safeguards used for tools like Ambient AI Scribe should be adapted for them; and fold AI-related risk monitoring into patient safety programs, with staff trained to identify and report AI-related safety events (VA OIG, 2026, p. 9). VA concurred with the second and third and concurred only “in principle” with the first. The target date for all three is April 2027 (pp. 11–13).

What the record does not show

The report does not document any patient who was harmed by an AI-drafted note, and it should not be read as saying one was. It also cannot tell us how often clinicians use these tools for clinical work, because VA itself cannot measure that. The engagement figures describe Teams sites, not patient encounters. The hallucination rates come from an outside study the inspector general cited, not from testing of VA’s own tools. And the inspector general’s findings, by its own statement, do not define a standard of care or establish legal liability (VA OIG, 2026, p. 10). What the record does show is a gap in classification and a gap in reporting, which together mean that if harm had happened, it would have been very hard to see.

Questions you can ask

• If you are a patient: you can ask your care team whether any AI tool helped write or summarize your notes, and who checked the result. If something in your record looks wrong, ask for it to be reviewed and corrected.
• If you are a clinician: ask whether the chat tool you use has been approved for clinical documentation, and how you would report a near miss in which it got something wrong.
• If you run AI in a health organization: list every tool staff use for clinical work, including general-purpose ones, and classify each by what it is used for, not what it was built for.
• If you run patient safety or quality: check whether your incident form has a way to record that AI contributed. If it does not, add one before you need it.

If AI scribes are on your mind, our earlier investigation of Ontario’s approved AI medical scribes looks at what testing found when someone did check.

Oversight reports like this one are rarely read by the people they affect. Reading them closely, explaining them plainly and asking what happens next is what FairByDesign does, and it only works if readers fund it.

---

Help keep independent scrutiny going. Veterans, patients and the public deserve someone who reads the oversight reports and asks what happens next. FairByDesign takes no money from the agencies or companies we write about, so your support is what keeps that line clear. Give once ($5, $10, $25, $50 or any amount), give monthly ($10 or $25), or become a patron: fairbydesign.org/patrons-2. Support never buys editorial control.

Not in a position to give? The book is free for everyone: read Operational AI Readiness.

References

Asgari, E., et al. (2025). A framework to assess clinical safety and hallucination rates of LLMs for medical text summarisation. npj Digital Medicine, 8, 274. https://doi.org/10.1038/s41746-025-01670-7 (as cited in VA Office of Inspector General, 2026)

Bakos, V. (2026). Operational AI readiness: A practitioner guide to governing AI that is already live. FairByDesign. https://fairbydesign.org/operational-ai-readiness/

Office of Management and Budget. (2025, April 3). Accelerating federal use of AI through innovation, governance, and public trust (Memorandum M-25-21). Executive Office of the President. https://www.whitehouse.gov/wp-content/uploads/2025/02/M-25-21-Accelerating-Federal-Use-of-AI-through-Innovation-Governance-and-Public-Trust.pdf

VA Office of Inspector General. (2026, June 11). Review of generative artificial intelligence chat tools for clinical use (Report No. 26-00182-140). U.S. Department of Veterans Affairs. https://www.vaoig.gov/reports/national-healthcare-review/review-generative-artificial-intelligence-chat-tools-clinical