U.S. FDA Launches Agency-Wide AI Tool “Elsa” to Optimize Performance

Launch & Purpose

• Elsa launched in June 2025 as a generative AI assistant intended to support FDA staff across divisions.
• Developed by Deloitte using Anthropic’s Claude, Elsa is designed for:
  • Clinical protocol review support
  • Summarizing adverse event reports
  • Drug label comparisons
  • Coding assistance for nonclinical databases
  • Identifying inspection targets
• Elsa is deployed agency-wide—from scientific reviewers to investigators—as part of the FDA’s larger AI modernization strategy.
• The tool operates within Amazon Web Services’ GovCloud, a secure federal infrastructure. While the FDA states Elsa avoids training on proprietary industry data, it hasn’t disclosed the full scope of its training datasets.

Aggressive Timeline & Claims

• FDA Commissioner Marty Makary launched Elsa ahead of his June 30 deadline, stating the agency is “ahead of schedule and under budget.”
• In an interview, Makary cited internal reports that Elsa helped a reviewer complete a two-day task in six minutes—a claim disputed by some staff.
• Chief AI Officer Jeremy Walsh declared Elsa represents “the dawn of the AI era at the FDA,” with future plans for expanded generative AI integration.
• The FDA has framed Elsa as a foundational technology for faster drug, food, medical device, and diagnostics review workflows.

Employee Concerns

• Internal feedback, reported by NBC, Axios, Stat, and TechTarget, reveals systemic issues:
  • Elsa provides incomplete or incorrect summaries, especially when asked about FDA-approved products or public-facing information.
  • It struggles with basic functionality like uploading documents and answering user-submitted queries.
  • It’s not integrated with internal FDA systems and cannot access real-time or paywalled content, limiting its utility.
• Staff at the Center for Devices and Radiological Health (CDRH) say Elsa’s predecessor, CDRH-GPT, is still in beta and faces identical limitations—raising concerns about system maturity.
• Many employees feel Elsa is only suitable for administrative support and not ready for regulatory or scientific tasks.
• There is no formal policy framework or usage oversight. Ethics experts, such as Richard Painter (University of Minnesota), have raised concerns about conflicts of interest, warning that AI-related contracts could compromise regulatory integrity.
• Some reviewers fear Elsa could eventually replace their roles. Layoffs and hiring freezes have exacerbated job security anxiety, particularly after the loss of 3,500 staff and a 25% proposed cut to the HHS budget.

Technical Background

• Since 2020, Deloitte has developed infrastructure for FDA’s AI tools:
  • Received $13.8 million for building the initial database of FDA documents.
  • In April 2025, was awarded $14.7 million to scale Elsa.
• Elsa evolved from CDER-GPT, originally built by the Center for Drug Evaluation and Research, which replaced CDRH-GPT after cost-cutting and was rebranded Elsa for wider rollout.

Security Context

• Hosted in GovCloud, Elsa ensures:
  • No internet connectivity
  • Secure internal access for FDA employees
  • Exclusion of proprietary submission data from training
• However, Axios and NBC report that the FDA has not disclosed training data details nor implemented robust AI-specific cybersecurity protocols.

Why This Matters

• Public Safety Risk: Flawed summaries or decision support from Elsa in clinical or drug review tasks could endanger health outcomes.
• Governance Deficit: Elsa launched without guardrails, oversight bodies, or ethical frameworks—concerning for a tool embedded in public health decision-making.
• Transparency Gaps: The AI’s training base, evaluation methodology, and technical documentation remain undisclosed, limiting accountability.
• Labor Displacement: AI integration amid layoffs and resource strain stokes fears of job replacement, eroding morale and institutional trust.
• Ethics Red Flags: Experts stress the need for independent safeguards to prevent conflicts of interest in AI contracting or tool usage in regulatory settings.
• Federal Trend Indicator: Elsa reflects a broader U.S. federal shift toward fast, top-down AI integration without adequate operational maturity or public safeguards.