FDA’s rushed AI tool rollout faces significant challenges

The FDA‘s hasty rollout of artificial intelligence tools is raising serious concerns among agency staff, who report that the new agency-wide AI system is providing inaccurate information despite leadership’s enthusiasm. This tension highlights a growing divide between the Trump administration’s aggressive AI implementation goals and the practical realities of deploying reliable AI systems in regulatory contexts where precision and accuracy are paramount.

The big picture: The FDA has prematurely launched an agency-wide large language model called Elsa, despite staff concerns about accuracy and functionality.

• Commissioner Marty Makary proudly announced the rollout was “ahead of schedule and under budget,” emphasizing speed over performance.
• The system is being promoted to enhance everything from clinical protocol reviews to identifying inspection targets and summarizing adverse events.

Behind the claims: FDA staffers testing Elsa discovered significant problems with the tool’s reliability and accuracy.

• When questioned about FDA-approved products or public information, Elsa provided summaries that were either completely or partially incorrect.
• Employees told Stat news the tool was “rushed” and its capabilities were “overinflated” by officials, including those at the Department of Government Efficiency headed by Elon Musk.

Key details: The AI system is based on Anthropic’s Claude LLM and is being developed by consulting firm Deloitte.

• Deloitte has received $13.8 million since 2020 to develop the original database of FDA documents that forms Elsa’s training data.
• In April, the firm was awarded an additional $14.7 million contract to scale the technology across the agency.

The development process: Elsa emerged from competing internal AI initiatives following recent cost-cutting measures.

• Initially, each FDA center was working on its own AI pilot program.
• After budget cuts in May, the AI pilot from the Center for Drug Evaluation and Research (CDER-GPT) was selected for agency-wide scaling and rebranded as Elsa.

What they’re saying: FDA Chief AI Officer Jeremy Walsh characterized the launch as transformative for the agency.

• “Today marks the dawn of the AI era at the FDA. With the release of Elsa, AI is no longer a distant promise but a dynamic force enhancing and optimizing the performance and potential of every employee,” Walsh stated.

Counterpoints: Staffers warn that the current version of Elsa is not ready for scientific applications.

• “Makary and DOGE think AI can replace staff and cut review times, but it decidedly cannot,” one employee told Stat news.
• FDA employees also noted the agency has failed to establish proper guardrails for the tool’s use, with one stating: “I’m not sure in their rush to get it out that anyone is thinking through policy and use.”

Technical challenges: Problems extend beyond Elsa to other AI initiatives within the agency.

• Staff at the Center for Devices and Radiological Health reported that their AI pilot, CDRH-GPT, is buggy and disconnected from both the internet and FDA’s internal systems.
• Users have encountered problems uploading documents and submitting questions to the system.