FDA Goes All In on AI: Will Policy Direction Shift?
The survival of artificial intelligence at the FDA is no longer in doubt, but with the departure of its biggest champion, former Commissioner Makary, the path of integration and evolution may no longer be linear. Acting Commissioner Diamantis says AI remains a priority, but the successive departures of key figures such as the Chief AI Officer have raised questions about implementation prospects. Industry experts point out that the FDA's internal AI governance structure remains unclear and may revert to a decentralized departmental model; meanwhile, AI policies for sponsors have not loosened. Based on multiple sources, this article analyzes the evolution of the FDA's internal AI tool Elsa, transparency gaps, and future policy balance challenges.
Artificial intelligence's presence at the FDA is now a certainty, but its integration and evolution path may no longer be a straight line—now that one of the technology's most powerful advocates, former Commissioner Dr. Marty Makary, has stepped down.
Makary had pushed for a more centralized approach to AI governance and launched a tool calledElsa, an agency-level tool. This large language model can handle supportive tasks such as drafting and summarizing reports, thereby accelerating reviews and assessments. In December 2024, the FDA also announced plans to expand the use of agentic AI inpremarket review, inspections, and administrative matters.
According to research from the Bipartisan Policy Center, the number of AI use cases reported by the FDA between 2024 and 2025surged by 148%, against the broader backdrop of a government-wide push to adopt the technology.
Acting Commissioner Kyle Diamantas said AI remains a top priority for the FDA. Even so, the recent string of high-level departures—includingMakary, Chief Artificial Intelligence OfficerJeremy Walsh, and Acting Chief Information Officer Sridhar Mantha—has raised questions about the prospects for AI implementation.
"It remains unclear how the leadership and governance structure for the FDA's overall AI efforts will evolve," said Tala Fakhouri, chief AI and regulatory strategy officer at Parexel and a former FDA AI policy official.
Fakhouri noted that when FDA officials recently discussed AI at conferences, they appeared as representatives of individual sub-offices, which may signal a return to the earlier, decentralized, siloed approach to AI advancement. Meanwhile, efforts to increase transparency around AI use and develop related policies may also slow down.
These potential setbacks mainly affect the FDA's internal use of AI, not the regulatory policies governing drug companies' own use of the technology. To date, the FDA's policies on sponsor use of AI have not changed.
"I don't expect changes there, which is good news," Fakhouri said.
The continued evolution of AI applications
In recent years, the FDA's internal use of AI to lighten employees' workloads has developed steadily. Fakhouri explained that Elsa originated from the CDER GPT project developed by CDER. The FDA expanded on that original project by adopting a retrieval-augmented generation (RAG) system to reduce AI hallucinations. The system is designed to confine the large language model to a defined database of trusted information and is customized for each center within the agency.
This allows employees to use the tool to access information relevant to their job responsibilities. For example, employees can quickly generate a summary of industry comments on a specific proposal, or the Office of New Drugs can use Elsa to quickly generate a regulatory submission history spanning several years.
Although Elsa has been used for such heavy-lifting tasks, Fakhouri believes the technology has not been involved in final decision-making.
However, "employees can use these tools to enhance the work they are doing. We should all be pleased about that," she said.
Where the FDA's AI program should go from here
Fakhouri said transparency regarding how the FDA uses AI in its processes remains insufficient.
"If regulators are using AI in specific ways to augment reviewer work or act as review assistants, I think it would be good practice for industry to understand the specific forms of these uses," Fakhouri said.
Detailing how the FDA uses AI could also help foster collaboration with industry.
"If I'm a sponsor or a CRO preparing submissions on behalf of a sponsor, I can have all the data labels and information ready so that reviewers and (AI assistants) can review my package more effectively," she said.
Fakhouri expects that, over time, the FDA will move toward greater transparency.
"But that requires someone in a leadership position within the agency to recognize it and truly drive it forward," she added.
Fakhouri also hopes the FDA will streamline AI-related rulemaking that affects industry, such as how AI tools used in clinical trials should be validated.
Under Makary's leadership, policy changes were sometimes announced through journal articles or press conferences rather than thetraditional FDA guidanceprocess. But under Diamantas, the FDA appears to be returning to established norms. Diamantasrecently confirmedthat informal statements made by the former commissioner do not represent official policy.
"They will follow the regular guidance and policy development process," Fakhouri said.
While the traditional rulemaking process shields the pharmaceutical industry from uncertainty, it can also be a slower process, posing its own challenges at a time when drug companies are rapidly adopting different AI platforms.
"Even the fastest guidance release could take a year. In the age of AI, a year is a very long time," she said.
Finding a balance between structure and flexibility will be one of the key policy challenges facing the FDA.
"I'm not sure what the solution is, but there does need to be more agile and flexible policymaking, and perhaps more frequent communication between industry and regulators," she said.