AI tools like Elsa help the FDA by summarizing reports and reducing paperwork, but about one in five of Elsa’s reports have fake or wrong information. This raises concerns about over-reliance and missed mistakes, though leaders are optimistic about fixes. Most staff use Elsa only for simple tasks, and the FDA is now transparently reporting errors to build trust and guide other agencies.
What are the main benefits and risks of using AI tools like Elsa in FDA regulatory review?
AI tools in FDA regulatory review benefits and risks such as Elsa can automate up to 18% of routine FDA paperwork, improving efficiency by summarizing reports and drafting code. However, about 20% of citations generated are inaccurate or fabricated, raising concerns about automation bias and reliability in regulatory reviews.
Across FDA offices the newest staff member is not a post-doc pharmacologist but a cloud-based assistant named Elsa. Rolled out in June 2025, the tool was pitched as a way to cut review times by summarizing adverse-event reports and drafting database code. Early metrics suggested up to 18% of routine paperwork could be automated, a figure that now competes for attention with a more sobering statistic: internal audits show that roughly one in five citations generated by Elsa either misrepresents an existing study or invents one outright.
Outside observers note that the episode fits a broader trend documented by applied-clinical-trialsonline.com: large language models in healthcare generate plausible but false outputs at rates between 15 % and 30 %, especially when asked to supply references. The same survey found that nearly half of regulatory professionals worry about “automation bias” – the human tendency to trust machine output without adequate skepticism.
Commissioner Dr. Marty Makary continues to champion AI adoption, arguing that iterative fixes will tame the hallucination problem. Yet internal surveys cited by RAPS show only 38 % of reviewers feel comfortable relying on Elsa for anything beyond first-pass literature searches. With other agencies watching closely, the FDA has promised monthly transparency reports on error rates, a move that may shape how the European Medicines Agency and Health Canada craft their own AI road maps.
