Dr. Marty Makary in an FDA video
The FDA has launched ‘Elsa,’ its new AWS GovCloud-based generative AI assistant. Commissioner Dr. Marty Makary reported the tool, rolled out “ahead of schedule,” is already cutting some scientific review tasks from “two to three days” down to “six minutes.”
This agency-wide deployment of Elsa, a secure in-house Large Language Model (LLM), was also achieved “under budget,” Dr. Makary confirmed. In a video accompanying the announcement, he stated, “[We] announced the goal date of June 30… today, we met that goal ahead of schedule and under budget thanks to the willingness and collaboration of our in-house scientific leaders across the centers.”
Makary said the real win is Elsa’s larger remit to “modernize how the agency functions.” He noted that Elsa is already being actively deployed across a range of critical agency tasks, with a focus on streamlining complex review processes, including to “expedite clinical protocol reviews and reduce the overall time to complete scientific reviews.” Some of the clearest improvements, he noted, are in the significant time savings for specific, labor-intensive tasks. Makary shared an anecdote: “One scientific reviewer told me what took him two to three days, now takes six minutes [with Elsa].”
While the FDA touts these early successes and rapid implementation, the initiative has also prompted early questions regarding its broader impact and the readiness for such swift, agency-wide adoption.
Broad acceptance across agency functions
Beyond these core review accelerations, Elsa is finding broader acceptance among FDA staff for a variety of duties. According to Makary,s some employees are already embracing the tool for diverse applications. “Some of the other ways staff have already told me they love using the AI tool include summarizing adverse events to support safety profile assessments, conducting expedited label comparisons, and generating code to facilitate the development of databases for non-clinical applications,” he said.
Elsa’s utility also extends to the agency’s field operations and oversight functions. “We’re also using AI to help inspectors identify high-yield inspection targets,” Makary said.
Industry observers note that AI adoption in regulatory environments was likely inevitable given widespread LLM usage across workplaces. “If you’re at the FDA, you can imagine that there inevitably will be users, people that work at the FDA, who are going to use an LLM for some of the work that they’re doing. So the best thing that the FDA could do is say, ‘Okay, we’re going to tune this [large language model Elsa] on materials that are more likely to be the use cases that we have at the FDA,'” said Ahmed Elsayyad, president and co-founder of Ostro.
Concerns about rushed deployment
While the FDA projects significant benefits from Elsa, its rapid, agency-wide deployment has prompted some consternation in some quarters. For instance, STAT News and POLITICO reported that some current and former FDA employees privately expressed concerns that the tool’s rollout may have been “rushed” and its capabilities “over-inflated,” with one staffer telling POLITICO, “We aren’t anywhere near this [shaving significant time off product review].” Similarly, Pharmaphorum wrote in its launch-day coverage that the FDA was moving at “breakneck” speed to roll out a gen-AI tool “amid staff cuts.” Meanwhile, the law firm Manatt, among others, have cited the potential of genAI models to hallucinate, or fabricate material, in some cases. Previously, the FDA’s own Digital Health Advisory Committee has cautioned that GenAI outputs must be shown to be “accurate, consistent and reliable” despite such risks. Still, there are methodologies available such as retrieval augmented generation, among others, that can limit such problems, though the agency hasn’t divulged details about the implementation.
Given the high-stakes nature of regulatory decision-making, experts emphasize the importance of human oversight in AI systems. “Having ‘human in the loop’ is going to be very important… having human in the loop to verify those citations and to ensure that the information—that a) these are valid citations, and b) they’re coming from reliable sources,” said Chase Feiger, MD, CEO and co-founder of Ostro.
Security and data protection challenges
Security is another flashpoint. The complexity of implementing advanced AI capabilities while maintaining data security presents ongoing challenges for regulatory agencies. “Would you want the FDA to be utilizing browser use to actually download files to make actions on your behalf based off of confidential information? Probably not,” Elsayyad of Ostro noted.
While Commissioner Makary insists all data stay inside GovCloud, POLITICO reports that some drug and device makers have already asked the agency to keep their submissions out of Elsa’s reach. On the broader point, Makary noted: “One question I’m frequently asked is about the security of the tool. So I want to be clear, all information stays within the agency, and the AI models are not being trained on data submitted by the industry.”
Staff morale and future plans
Morale is another consideration. Even before Elsa’s launch, RAPS reported that “FDA staff are more concerned about meeting user fee review goals and completing essential tasks, though some are anxious about losing their jobs and have begun looking for opportunities in the private sector” owing to reports of impending layoffs and buyout offers.
Feiger of Ostro commented on this general principle, emphasizing the necessity of controlled AI systems in regulatory settings. “Regulators can’t let reviewers cut-and-paste confidential sections into public chatbots,” Feiger noted. “A locked-down, internal model like Elsa is the only logical fix to leverage AI’s power securely.”
Commissioner Makary positioned Elsa’s rollout not as a final destination, but as a foundational element of the FDA’s evolving AI strategy. He signaled a commitment to continued technological advancement within the agency. “Today’s rollout of Elsa will be the first of many initiatives to come as the FDA taps AI to rapidly transform how we do business to better serve you,” Makary concluded.
