EPA Administrator Lee Zeldin has vowed to expedite chemical reviews, and he’s counting on yet-to-be-developed artificial intelligence models to do so.
While experts say the core technology already exists, they caution that the agency still faces significant hurdles in data quality and trust.
EPA is eyeing development of an “AI Chemist Assistant” that “will help chemical reviewers search various repositories to identify chemical and chemical analog information used in TSCA [Toxic Substances Control Act] submission reviews and risk evaluations, possibly saving hundreds of staff hours per review/evaluation,” according to the agency’s internal AI use case inventory, reviewed by POLITICO’s E&E News.
Also listed on the internal AI use case inventory is a tool called “EcoVault,” which is intended “to summarize and capture key information from scientific studies and other lengthy unstructured documents to provide narrative summaries and interaction abilities to OCSPP [Office of Chemical Safety and Pollution Prevention] staff.”
The chemical industry has long complained about delayed review processes impeding innovation, as the agency struggles to meet its up-to-180-day deadline for new chemical assessments under TSCA.
Zeldin has said speeding up new chemical reviews is a priority. He told the House lawmakers during a hearing in May that he was “confident that we are going to be able to tackle the backlogs that we inherited” in part by replacing “outdated technology” with “new technology, including AI.”
EPA press secretary Carolyn Holran confirmed in an email that the “chemical safety program is going to support Administrator Zeldin’s focus on artificial intelligence by using AI tools and machine learning algorithms to transform and improve the way we review scientific information — making our chemical safety reviews more efficient and robust.”
“We’re evaluating opportunities to use AI to streamline our new chemical reviews and to inform our risk evaluations to ensure we’re using the right technology to support our program,” Holran continued.
The push aligns with a broader surge in EPA’s adoption of AI under the current administration. The agency’s internal AI use case inventory had documented 103 projects as of August, an increase from just 17 such cases in late 2024.
Thomas Hartung, a professor in the Johns Hopkins Bloomberg School of Public Health, has spent more than a decade researching the ways AI technologies can transform toxicology.
“Toxicology used to be an extremely data poor field,” Hartung said. “But because of all of the progress in recent years from high content imaging or mixed technologies, it has become a data rich discipline.”
Ensuring datasets are robust, open and of the highest quality is the first step when it comes to AI readiness, Hartung said.
Niki Maslin, EPA’s chief technology officer and chief artificial intelligence officer, echoed a similar sentiment.
“The tool is only as good as the data that’s going in,” Maslin said in a recent interview. “So we want good data going into the process, and that way we can ensure the answers that come out of the tools that we’re using have the best chance of being as accurate as possible.”
Bhavik Bakshi, a chemical engineering professor at Arizona State University, cautioned that major data gaps still exist.
“The problem is nobody has found everything yet, nobody has compiled it yet, in the way that we have been doing over the last five to six years,” Bakshi said.
Bakshi and his team were developing an open-access model of the U.S. chemicals industry funded by EPA grant money, but that grant was canceled not long after the first installments were disbursed.
While Bakshi’s project was focused on data that was publicly available, Hartung said another concern is use of confidential business information, which can be toxicology data manufacturers submit to EPA for review but are protected as industry trade secrets.
“This data are the data of the companies,” Hartung said. “They’re not owned by us and not owned by you.”
That’s raises concerns when it comes to reproducing scientific findings, which the Trump administration has doubled down on in its “gold-standard” scientific integrity policy.
“AI should be much better in explaining why it is coming to certain conclusions,” Hartung said.
EPA’s current AI policy prohibits staff from relying on AI-generated answers without thorough verification, specifically to prevent the kind of “hallucinations” where models confidently make up false information.
Telly Lovelace, spokesperson for the industry trade group American Chemistry Council, welcomed the agency’s “continued exploration” of AI uses, touting “the potential to significantly reduce review times, improve consistency, and alleviate resource burdens.”
Lovelace also echoed the importance of “human review and oversight,” safeguards “to protect confidential business information” and that “AI model inputs and outputs should be sufficiently transparent to facilitate external validation and reproducibility when used for government activities.”
Tamara Kneese, the director of Data & Society’s climate, technology and justice program, said she is skeptical of the Trump administration’s accelerated implementation of AI being done in a thoughtful way.
“I think there is a weird fantasy that AI can do science the same way … but I think it’s also more cynical than that, I think it is also just trying to destroy the entire scientific research channel and research apparatus in the U.S.,” Kneese said. “I think what they really mean, honestly, is that they’re just going to try to reduce environmental protections, including impact assessments.”
Hartung said that even though the methods to develop systems to do chemical risk assessments are here or emerging, “it is really about management and trust. That’s what it is.”
“I would say that in a certain phase of evaluation, especially now in the phase of building trust, it needs to be used by specialists. That’s the validation exercise,” Hartung continued.
Reach reporter Ellie Borst on Signal at eborst.64.
