In recent weeks, senior executives from OpenAI have held multiple meetings with the U.S. Food and Drug Administration (FDA) to explore how artificial intelligence could be used in the drug approval process. Central to these discussions is a project called cderGPT, believed to reference the FDA’s Center for Drug Evaluation and Research.

According to sources familiar with the matter, these meetings reflect a broader initiative within the FDA to modernize and accelerate drug evaluations using advanced AI technologies.

Pushing for AI-Driven Reform

FDA Commissioner Marty Makary recently voiced frustration with the lengthy timeline of drug approvals. “Why does it take over 10 years for a new drug to come to market?” he posted on X. “Why are we not modernized with AI and other things?” Makary revealed that the FDA has already completed its first AI-assisted scientific review and suggested it’s only the beginning.

While Makary didn’t explicitly mention OpenAI, insiders confirm that a small team from the company has met several times with FDA representatives and two associates of the so-called Department of Government Efficiency, linked to Elon Musk. The focus of these conversations has been the development of cderGPT, likely aimed at AI applications for drug evaluation. Leading these talks on the FDA side is Jeremy Walsh, the agency’s newly appointed AI officer. No formal contracts have been signed yet.

Broader Government AI Collaboration

Walsh has also been in talks with Peter Bowman-Davis, currently acting chief AI officer at the Department of Health and Human Services. Bowman-Davis, a Yale undergraduate on academic leave, is affiliated with Andreessen Horowitz’s American Dynamism team.

Meanwhile, former FDA Commissioner Robert Califf confirmed that AI has been in use at the agency for several years, though he noted the importance of transparency around what “AI-assisted” review actually entails. “There has always been a quest to shorten review times,” he said, adding that AI could play a significant role.

The Scope and Limits of AI in Drug Review

Though AI could help streamline portions of the drug approval pipeline, experts caution that it would only shave off a small part of a much longer and failure-prone development cycle. Most experimental drugs never make it to the FDA review stage at all.

Rafael Rosengarten, CEO of Genialis and a cofounder of the Alliance for AI in Healthcare, supports using AI for certain repetitive tasks, like ensuring an application is complete. But he stresses the need for policy frameworks around data quality and model performance. “These machines are incredibly adept at learning,” he says, “but they need to be trained properly.”

More advanced AI use in regulatory decisions would need thorough development, testing, and validation before becoming standard practice.

A former FDA staffer who has experimented with ChatGPT in clinical contexts raised concerns about the risk of “hallucinated” responses. “The ability of these models to sound convincing, even when wrong, makes it unclear how reliable they’d be for reviewers.”

AI and Regulatory Infrastructure

Currently, FDA reviews typically take about a year. However, several accelerated pathways already exist for breakthrough and high-priority drugs, including the Fast Track and Breakthrough Therapy designations.

Andrew Powaleny, a spokesperson for the pharmaceutical lobbying group PhRMA, emphasized that while AI holds promise, implementation must be “thoughtful and risk-based, with patients at the center.”

The FDA itself has been investing in AI internally. In December 2023, it launched a fellowship program focused on developing large language models for use in precision medicine and regulatory science.

Meanwhile, OpenAI has also been preparing its tools for government adoption. In January, it unveiled ChatGPT Gov, a self-hosted version of its AI chatbot tailored for regulatory compliance. The company is working to obtain FedRAMP moderate and high accreditations for ChatGPT Enterprise—certifications necessary to handle sensitive government data.

Source