A watchdog inside the Department of Veterans Affairs issued an urgent advisory Thursday over two AI chat tools currently in use by VA healthcare providers, citing “potential patient safety risks.”
Doctors aren’t using AI to diagnose patients, but the Inspector General found problems could potentially arise in some cases when doctors use artificial intelligence to analyze medical information and update patient records.
The bulletin said the systems can be vulnerable to producing misinformation, privacy violations, and bias, and that the systems had been put in place without review by the VA’s own patient safety experts.
When reached for comment about the advisory, Pete Kasperowicz, a spokesperson for the department, said that “VA clinicians only use AI as a support tool, and decisions about patient care are always made by the appropriate VA staff.”
Dr. Matthew Miller, the former executive director for VA Suicide Prevention, told Task & Purpose that the memo was obviously urgent.
“It’s an official communication from [the Office of the Inspector General] recommending a ‘stopping of the presses’ of the AI tools under review based upon perceived patient safety concerns,” Miller said. “Specifically, [the Office of the Inspector General] has come across info within their review that indicates appropriate process-oriented safeguards may not be in place.”
Dr. David Shulkin, who served as VA Secretary in 2017 and 2018 and now does technology consulting for healthcare companies, said he doesn’t think the report should be seen as an “argument for VA to stop or slow down what it’s doing.”
“I actually think that it is important that veterans know that VA has the best and the most capable technologies for them to get their care, so I don’t think that you want this to be a reason why VA steers away from artificial intelligence,” he said. “But I do think that this was an appropriate report to suggest that there’s more work to be done here, to assure that the technologies are being used appropriately, and that we maintain patient safety as the highest priority for the veterans.”
Two chatbots, but little oversight
The advisory was released as a Preliminary Result Advisory Memorandum or PRAM, is only two pages long and did not include specific cases where patient safety was put at risk.
According to the memo, VA health providers can now access two AI chat-style systems for use with patient care, VA GPT and Microsoft 365 Copilot Chat. The tools, which are both Large Language Models, or LLM-style chat systems, are intended to help simplify and reduce time spent on documenting patient care. Doctors and medical providers enter patients’ clinical information into the chat tool, which can be copied into veterans’ electronic health records and “can be used to support medical decision-making.”
But, the IG cited research that found that AI systems can introduce errors or omit correct data in somewhere between 1% and 3% of cases using patient medical data. Federal investigators stated that in the event that generative AI tools’ results in incorrect or omitted data, it “may affect diagnosis and treatment decisions,” referencing a May 2025 study on the risks of LLM’s used for medical documentation. The study was done by researchers with TORTUS AI, an AI assistant tool used by clinicians, and Guy’s and St Thomas NHS Trust, a foundation under the UK’s National Health Service.
Top Stories This Week
According to the research, LLMs used for summarizing medical documents can cause errors referred to as “hallucinations” — creating information that was not part of the original data or completely omitting information from the original user’s input.
“Errors in clinical documentation generation can lead to inaccurate recording and communication of facts. Inaccuracies in the document summarisation task can introduce misleading details into transcribed conversations or summaries, potentially delaying diagnoses and causing unnecessary patient anxiety,” the researchers wrote.
A goal of reduced ‘burnout’
The upside, researchers found, is that AI might help clinicians who spend a great amount of their time on paperwork, which raises the “cognitive load” on doctors and can “lead to burnout.” Other research has noted similar benefits of AI use for giving providers back time they can spend with patients as well as using AI tools to translate complex medical language into more accessible instructions for patients.
VA GPT, which was released by the agency as a generative AI pilot tool, had nearly 100,000 users in September 2025 and was estimated to save them two to three hours per week.
Through interviews, the IG found that the VA did not work with the agency’s National Center for Patient Safety before fielding these AI chat tools, nor does it have a formal mechanism in place to identify or fix risks that come with using generative AI tools.
Task & Purpose inquiries to the patient safety center were not immediately returned.
“Not having a process precludes a feedback loop and a means to detect patterns that could improve the safety and quality of AI chat tools used in clinical settings,” the advisory stated.
James McCormick, executive director of government affairs for Vietnam Veterans of America, said their group has not received complaints or problems regarding AI resources in the VA from their members.
“However, if issues have come up that could put members at risk, we certainly support a deeper dive into these tools for corrective action and improvements to ensure only the best quality of care and services for our veterans,” McCormick said.
