OpenAI is reportedly in talks with the U.S. Food and Drug Administration about using AI to help evaluate pharmaceuticals. The project at the center of those discussions is called "cderGPT," a name that appears connected to the FDA's Center for Drug Evaluation and Research.
The discussions are still preliminary. According to the source report, several meetings have taken place, but no formal agreement has been reached.
What cderGPT is reportedly meant to address
The reported project points to a practical question now facing regulators: where can large language models help inside complex review processes, and where would their limits create unacceptable uncertainty?
The FDA's work around pharmaceuticals involves scientific, medical and regulatory judgment. The source does not say which specific tasks cderGPT would perform, and that distinction matters. An AI system used for internal analysis would raise different questions than one used to influence regulatory conclusions.
What is clear from the report is that the idea is not appearing in isolation. The FDA has already been experimenting with AI for several years, according to former FDA Commissioner Robert Califf. He also noted that it remains unclear which parts of the process actually rely on AI.
That ambiguity is important. If an agency is testing large language models, readers need to know whether the technology is being used for summarization, search, drafting, internal support or more consequential review steps. The source article does not settle that point, so the most accurate reading is cautious: cderGPT is a discussed project, not a defined public system.
Who has been involved in the discussions
The meetings were reportedly led by Jeremy Walsh, the FDA's first AI officer. A small team from OpenAI also took part, along with representatives connected to Elon Musk's "Department of Government Efficiency."
Walsh also met with Peter Bowman-Davis, a Yale student currently on leave who now serves as acting Chief AI Officer at the U.S. Department of Health and Human Services. Bowman-Davis is part of the "American Dynamism" team at Andreessen Horowitz.
OpenAI declined to comment on the discussions. That leaves the current status narrow but significant: the company is reportedly talking with the FDA, the project has a name, meetings have occurred, and the agency has not reached a formal deal.
The involvement of officials focused on AI suggests that the FDA is treating the subject as an institutional technology question, not merely as a software procurement issue. A model built for regulatory use would have to fit into existing agency responsibilities and standards, especially if it handled sensitive government data or supported internal regulatory science work.
The FDA has already explored AI internally
The reported cderGPT talks follow earlier FDA activity around artificial intelligence. In December 2023, the agency issued a research grant focused on developing large language models for internal applications.
The source identifies precision medicine and regulatory science as examples of those internal applications. That gives the cderGPT discussion a broader context: the FDA has already been looking at how language models might support specialized work inside the agency.
OpenAI has also made a separate move toward government use. In January, the company introduced "ChatGPT Gov," a special government version of ChatGPT designed to meet U.S. regulatory requirements. The company is also seeking FedRAMP certification to handle sensitive government data.
Those details matter because pharmaceutical evaluation is not a casual use case. Any AI system in that setting would have to be judged not only by whether it is convenient, but by whether its training data, performance standards and failure modes are understood clearly enough for government use.
Support comes with demands for clear standards
Industry leaders and experts such as Rafael Rosengarten of Genialis generally support using AI to automate certain regulatory review steps. But that support comes with conditions.
The source highlights two areas where clear rules are being sought:
- Training data: What information is used to build or adapt a model for regulatory work matters because it shapes what the system can and cannot reliably handle.
- Performance standards: A model used in a regulatory setting needs measurable expectations, especially if it is assisting with review steps rather than simply providing general productivity support.
The central concern is familiar but serious: large language models can produce convincing hallucinations. In a normal office workflow, that may create a correction burden. In pharmaceutical evaluation, an output that sounds authoritative but is wrong could be much harder to tolerate.
That does not mean AI has no place in the process. The source notes general support for automating certain regulatory review steps. It does mean the boundary between assistance and reliance has to be explicit. A system that helps organize information is different from one that becomes part of the evaluative chain.
Why the cderGPT talks matter
The FDA already uses fast-track programs such as "Fast Track" and "Breakthrough Therapy" to accelerate drug reviews. The cderGPT discussions add another possible route for speeding or supporting work: AI-assisted internal processes.
But speed is only one side of the issue. The stronger question is whether AI can be introduced in a way that keeps the review process transparent, testable and accountable. The source does not describe a completed product, a launch plan or an agreed deployment. It describes talks, meetings and unresolved questions.
For now, cderGPT should be understood as a signal. OpenAI is reportedly exploring how its technology might fit into one of the most sensitive areas of government science administration, while the FDA has already been examining large language models for internal use.
The next meaningful development would be clarity: what the system would do, what data it would use, how its performance would be measured and whether it would remain a support tool or become part of regulatory review itself. Until then, the story is less about a finished AI system and more about the standards that would need to exist before one could be trusted.