The Food and Drug Administration (FDA) recently announced the launch of its new AI tool, Elsa, to help them review regulatory submissions. This rollout initially faced skepticism among scientists and regulatory experts about the potential for inaccuracies, AI hallucinations, transparency concerns, and other concerns. Even FDA staff had mixed reviews.

“When they first announced the tool, it did seem like a very aggressive rollout,” said Brynn Stanley, an attorney at Gardner Law, a law firm that advises FDA-regulated companies with regulatory submissions. “There wasn’t a lot of information that I could find as far as how the tool’s been validated and how exactly it’s being used.”

Elsa was designed to save the FDA time in reviewing regulatory submission by reading, writing, and summarizing documents. It can also write code to create databases and generate tables. Brandon Rice, Chief Product Officer at Weave Bio, a company that uses large language models to automate investigational new drug submission preparation, thinks that Elsa can be particularly helpful to transform narrative-style writing, which are easier for sponsors to write, into tables, which are easier for reviewers to understand.

So far the FDA hasn’t provided any guidance on how sponsors should tweak their submissions given AI-based initial review, but the industry is already responding and seeing the effects of the rollout.

One question that Steven Bulera, Corporate Vice President, Chief Scientific Officer: Discovery & Safety Assessment at Charles River Laboratories, has raised is if Elsa could “flag findings that a human would accept as nondrug-related.” He explained, “we're starting to get questions on background findings from pharma companies after their regulatory submissions.” These background findings are those found in animal studies that are considered normally occurring rather than an effect of the drug tested. If Elsa flags these nondrug-related findings, it could add more time to the review process because pharma companies would need to obtain and provide data on these findings to formulate a response to the FDA, Bulera said. He advises that if more questions are being asked about nondrug related findings, study directors will need to be more proactive in addressing these nondrug-related findings in their study reports.

When it comes to addressing AI inaccuracies, Rice pointed out that it’s important to make sure submission documents don’t have any ambiguities in language. “These models, as impressive as they are, are not thinking,” said Rice. “They’re just transforming information from one form into another.” He added, “if there are ambiguities in the source information, that can trip up the models and confuse them and then cause the output to be incorrect.” Stanley said that this advice is similar to what needs to be done for a human reviewer, but AI may not be able to account for the ambiguities in language that humans can. “It’s even more critical now,” she said.

While the industry has seen the rise of AI and the accompanying skepticism of the technology, the FDA’s adoption of Elsa only signals the start of the agency’s more widespread use of AI. At its current state, however, humans still play a role. Bulera said, “the final fact-checking and decision making is still going to need a human.”