A Search Engine That Writes Your Notes
VA gives staff two general-purpose chatbots approved for use with patient health information: VA GPT, which VA built, and Microsoft 365 Copilot Chat. Clinicians paste in clinical information along with a prompt, and the chatbot drafts a result.
How widely this happens, the OIG could not say. It found no means of measuring clinical use.
What it could see was a sample. Staff had uploaded 135 prompts to an internal sharing app, and 79 were clinical: 56 for drafting clinical notes, 17 for summarizing patient care, and 6 for other purposes.
VA treats a different tool, Ambient AI Scribe, as high-impact AI. That tool listens to visits and drafts notes, and the label triggers pre-deployment testing, an impact assessment, ongoing monitoring, and human oversight. VA did not give the chatbots the same label, though most of their shared clinical prompts did a similar job.
VA leaders told the OIG the chatbots are analogous to search engines, and they emphasized user-level responsibility. The OIG called the analogy flawed. A search engine finds links, it said, while these tools synthesize and transform sources to produce new content, such as a visit note drafted from a transcript.
The OIG cited a published study, not VA data, that found a 1.5 percent hallucination rate in AI-generated clinical documentation, with 44 percent of those errors considered major. The same study found a 3.5 percent omission rate, with 17 percent of omissions major.
VA’s own 2023 guidance warned that generative AI brings new risks and unknown consequences that could significantly harm veterans’ privacy and safety.
No Mentions in the Safety Reports
VA relies on the Joint Patient Safety Reporting system, which it shares with the Defense Health Agency, to capture errors and near misses. The system has no AI-specific category.
The OIG searched it for any mention of AI and found none. The system receives about 180,000 submissions a year.
That is not reassurance. A patient safety manager may file an incident under a different category based on the outcome and never record that a chatbot was involved. The center’s own AI lead put it plainly: “It’s very tricky. We don’t have the best answer yet.”
VA also cannot identify records created with AI after the fact, so it cannot go back and look for patterns.
The structure around the tools is thin. VHA’s National AI Institute and the Office of Information and Technology’s Chief AI Officer team coordinated in 2025 without a formalized governance relationship. VHA program offices and field staff can pursue their own AI development and purchases without formal approval.
The National Center for Patient Safety said it had one meeting with the Digital Health Office on AI. Its attempts to connect did not produce a joint work group.
In 2025, the center asked the Defense Health Agency to add AI to the reporting system’s list of causes, and no formal decision followed. Only after the OIG’s January advisory memo did VA move to add an AI category.
Minimize Unnecessary Bureaucratic Oversight
The OIG’s explanation is a policy choice. VA AI leaders said they prioritized rapid access to the chatbots over safeguards they deemed impractical.
One leader described the guidance as being to minimize unnecessary levels of bureaucratic oversight and push risk management down to the lowest reasonable level.
That view tracks OMB’s April 2025 memo on federal AI use, which directed agencies to remove unnecessary bureaucratic requirements that inhibit innovation. The same memo requires risk management proportionate to anticipated risks, with minimum practices for high-impact uses.
The OIG concluded that leaders’ emphasis on reducing barriers to innovation influenced oversight decisions, and that the viewpoint does not acknowledge how staff actually use the tools. FedScoop covered that finding in June.
One number from the coverage deserves a correction. Some outlets, including HIPAA Journal, reported that more than 15,000 staff were using the chatbots. That figure is not in the report.
The OIG said VA cannot measure clinical use. It offered two AI discussion sites on Microsoft Teams as a rough indicator of interest, with 10,997 and 4,835 active users over 90 days. The 15,000 figure appears to add those together, which is our inference, and those are people active in discussion channels, not users in patient care.
Who Owns April 2027
VHA concurred with all three recommendations, one in principle, and set April 2027 as the target for each. The plan covers guidance on permissible clinical uses, an evaluation of whether high-impact safeguards should apply, and AI risk monitoring inside patient safety programs.
The guidance will include penalties for inappropriate use of the tools.
The people attached to the plan have changed. VA’s chief technology and AI officer announced his departure in March, before the report was published.
The under secretary who signed VHA’s May 19 response resigned effective July 6. Acting officials took over both jobs.
The OIG used the same family of tool in this review. Its methodology appendix says the team used peer-reviewed, standardized, structured, and evaluated prompts in Copilot Chat to review inspection data such as interview transcripts, documents, and questionnaire responses.
It says the team then confirmed the output against the source material, edited the report, and took full responsibility for the content.
The appendix adds that the OIG’s use is not high-impact under the OMB memo, because its teams do not use AI as the principal basis for decisions or actions.
The inference here is ours. Both the OIG and VA reasoned from how a tool is used to whether it needs safeguards, and both rely on the user to verify the output.
The difference the report draws is curation. The OIG says its prompts were standardized and evaluated, while VA does not centrally curate or evaluate clinical prompts. The basis is the report’s methodology appendix and its findings.
The disclosure is welcome. It also leaves the question the report asked of VA: if a tool gets it wrong, how would anyone know?
