Insights / Regulatory strategy
It’s an AI vs AI World in FDA Review
FDA is now using even more sophisticated AI to review regulatory filings, and that impacts clinical trials and commercial strategy.
- Published
- Reading time 6 minutes
- Topics Regulatory strategy, AI leadership
FDA is now using even more sophisticated AI to review regulatory filings, and that impacts clinical trials and commercial strategy. In June 2025, the United States Food and Drug Administration (FDA) launched an innovative AI tool that promised to revolutionize regulatory review, FDA ELSA. The launch announcement itself read “AI is no longer a distant promise but a dynamic force enhancing and optimizing the performance and potential of every employee,” which then-FDA Chief AI Officer Jeremy Walsh said (notably, mere months ago in Q3 2026 he moved over to a Director role at Google DeepMind).
Meet ELSA, the New Superhuman Adversarial Reviewer on your FDA Filings
FDA’s ELSA launched June 2, 2025 running a stack using Claude (secondary reporting pings it as Sonnet 3.5 orchestrated in an AWS GovCloud environment, per Law360 and STAT News who claimed the inside track in chatting with FDA employees at the time). In December 2025, the FDA expanded from generative AI to modern agentic architecture that supports multi-step agentic AI workflows, building on ELSA, which the Agency itself claimed was already in use by more than 70% of its staff.
Early in 2026 this is where it started to get fun. Industry started to feel the impact of ELSA. Reviewer feedback on clinical trial applications, or Investigational New Drug Applications (IND) for drug trials and Investigational Device Exemptions (IDE) for device trials, as well as safety reports and packages for medicine and device approvals started getting feedback that was clearly AI generated. However, these were not clear “ELSA indicates” labels on reviewer feedback. They were data analysis and comments embedded in reviewer comments. ELSA was now flying (often invisibly to industry) as the co-pilot for regulatory findings. And soon we saw the same turn to audits, inspections where reviewers came prepared at the data table level across tens of thousands of documents—a feat only achievable with ELSA in the copilot seat.
Almost overnight, agentic review for regulatory filings and audits on the industry side became the new form of defense. “Is my Claude (or other AI) setup as good or better than FDA’s?” experts asked. Can my AI copilot identify and help me to correct issues before FDA’s ELSA does? Such rigor is not new in regulated fields, but the sheer ability of AI systems to look across broad bodies of documentation posed new challenges. But it also has blind spots, and those blind spots are model specific and often shift with each new release.
Like every AI technology, especially those of 2024 and 2025 early launches, many also criticized the ELSA AI tool for confabulating data and hallucinating citations. One of the known failures, which still occurs but was more common in early models was hallucinations of citations, mixing feasible-at-glance author names and journal names and issues, which appear reasonable but turn out not to exist. Researchers have studied this phenomenon, with data up to 2025 that systematically analyzed scientific papers showing that across 2.5 million papers (111 million citations), over 146 thousand citations in 2025 alone were possibly hallucinated (see also the separate Lancet audit reported by STAT). This begs the question, what limitations does AI have when reviewing, or in industry’s case often generating, safety relevant data?
FDA’s Transition from Claude to Google’s Gemini: Politics, Institutions, and AI
In February of 2026, a federal directive was issued that all executive agencies move off of Anthropic (the maker of Claude) platforms. What’s interesting here is that this was not a typical software contract transition or upgrade. Rather this came as a directive direct from the President following a Pentagon dispute with Anthropic leadership on terms of use for Claude in mass surveillance and other autonomous military applications.
Though in July 2025 the DOD had awarded contracts to Anthropic, Google, OpenAI, and xAI of up to $200 million each “to accelerate Department of Defense (DoD) adoption of advanced AI capabilities to address critical national security challenges,” but that somewhat equitable spend across the Frontier shifted in 2026. In March 2026 internal notices were sent to FDA and Department of Health and Human Services (HHS) employees and staff to let them know that Claude access was being discontinued and that Gemini would be used to continue as the model powering FDA’s ELSA program. For the foreseeable future, FDA and HHS appear to be building primarily on Google’s AI platform.
ELSA 4.0 and the Future of AI Review for Medicines: HALO and FDA’s new Approach to Data
The next major evolution of the ELSA tool came in May 2026, when FDA announced ELSA 4.0+HALO (Harmonized AI and Lifecycle Operations for Data) operating in a Google-backed architecture. The driving principle is shifting away from individual “copilot” or task-based assistance into more sophisticated integrated data querying and analysis, and underlying that is the HALO. HALO is FDA’s new data strategy to create a platform that brings together 40+ previously standalone FDA systems into a single platform, creating an environment where AI-assisted reviewers have instantaneous and rigorous access to FDA data.
Which AI Should I Use To Build a Regulatory Submission?
That leaves many researchers looking for an answer to one big question: which AI should I be using to build and de-risk my regulatory submissions? Some have suggested using all of the above, ChatGPT, Claude, and Gemini each have different strengths in preparing regulatory submissions. While qualitative, I am finding many experts and firms posting their own ‘which AI to use / not use’ pieces. The right answer is often the simple one. Use the tool that your team is most comfortable with, and build context environments and access to data and source verification into how you are using it. Some tips:
- Choosing an AI Provider: Frontier AI like Claude, ChatGPT, and Gemini all have pros and cons but strengths are in how they are used and integrated into your work environment. Have your teams demo each, understand their capabilities, and work with an AI implementation expert to understand how they fit your existing workflows.
- Consider Data Retention Needs: If working with sensitive information, explore enterprise plans and which providers and models have zero data retention, and if your organization qualifies.
- Data Architecture, Skills, Context: To get the best performance from organizational AI systems, such as those serving regulatory teams, context is king. That doesn’t always mean having a huge data lake (though it may for larger organizations). For smaller firms and teams, that means building smart context where preferred websites (such as FDA/HHS databases, clinicaltrials.gov, and others) are built into the AI Agent’s preferred sources right in the context.
- Giving the AI a Clear Role and Onboarding: Every time you open an AI-driven application, essentially the “intelligent” AI model is being onboarded to your project the first time. Think like an organizational manager, what would you want a smart senior expert joining your project to know? Be sure to provide that information, as files and instructions. It is also important to make the AI’s role clear (assistant, co-researcher, auditor etc.) so the AI can align to the job you need it to perform.
While this applies to the frontier, there are many other 3rd party specialized AI solution providers, and documentation, evaluation and re-evaluation, and enforced model API disclosure policies are critical when evaluating these third parties.
Is your Team Ready for AI-Enabled Authoring and FDA Review?
Every investment in AI compounds. When your AI is set up incorrectly, everyone on your team is held back. Conversely, a proper regulatory authoring context can perform on the level of FDA’s ELSA, and even in other targeted applications specific to your regulatory documentation, manufacturing and clinical trial needs.