VA Deployed AI Chat Tools Without Safety Oversight, Watchdog Finds

The Veterans Affairs Department rolled out generative artificial intelligence chat tools to clinical staff without establishing adequate oversight or safety protocols, according to a report released Thursday by the VA's Office of the Inspector General.

The watchdog found that despite active use of AI tools throughout the Veterans Health Administration, the absence of proper governance structures created "risks for patient safety" and hampered the department's ability to monitor for errors. The OIG conducted its review between October 2025 and February 2026, according to details first reported by FedScoop.

Tools deployed without uniform safeguards

VA clinical staff currently use VA GPT and Copilot Chat for various tasks. The department is also piloting Ambient AI Scribe for drafting clinical notes, which the report noted has more robust governance mechanisms in place than the chat tools.

The disparity in oversight troubled investigators. The OIG expressed concern about VHA's capacity to protect patient safety when chat tools are deployed in clinical settings without the same protocols implemented for Ambient AI Scribe—specifically feedback loops and pattern-detection capabilities that could improve clinical use and prevent future safety incidents.

Push to reduce bureaucracy created vulnerabilities

The report found that VA's efforts to eliminate "unnecessary levels of bureaucratic oversight"—aligned with a 2025 Office of Management and Budget memorandum on AI—may have introduced new vulnerabilities rather than simply streamlining operations.

"VA leaders' emphasis on reducing barriers to innovation influenced oversight decisions," the report stated. The OIG warned this approach fails to acknowledge the high-impact nature of AI chat tools used by VA staff, resulting in inadequate risk management practices despite the same OMB memo's requirements for addressing "anticipated risks from [AI's] use."

Interviews with VA AI leaders revealed they prioritized rapid access to AI chat tools over safeguards they considered impractical. These leaders had not effectively coordinated with VHA's National Center for Patient Safety and compared generative AI to a search engine—a comparison the OIG deemed flawed because generative AI synthesizes information rather than simply retrieving it.

Critical infrastructure gaps identified

The VA lacks processes and infrastructure to ensure AI patient safety communication and monitoring, the report found. The department has no established methods to identify records created with AI assistance, which would be necessary to retroactively investigate safety concerns.

The OIG issued recommendations for the VA to address generative AI chat tool governance, evaluate AI chat tools as high-impact systems, require appropriate safeguards, and integrate monitoring of AI-related risks into existing patient safety programs.

Why it matters

This report illustrates a fundamental tension in federal AI adoption: the pressure to innovate quickly versus the imperative to maintain safety standards in high-stakes environments. For healthcare organizations deploying AI, the VA's experience demonstrates that treating generative AI tools as low-risk productivity enhancers—rather than clinical decision-support systems requiring rigorous oversight—can create systemic vulnerabilities. The findings suggest that deregulatory approaches to AI governance may be premature when applied to domains where errors carry serious consequences.

In response to the draft report, the VA concurred with all recommendations and outlined steps to increase communication with AI governance bodies, health agencies, and reporting systems. The department provided action plans to develop clinical AI governance and education programs.

"VHA greatly values the OIG's assistance in ensuring that all stakeholders are unified in supporting VHA's vision of providing all Veterans with access to the highest quality care," said VA Under Secretary for Health John J. Bartrum.

FedScoop first reported these findings.