The EPA may soon deploy AI to regulate toxic chemicals. It's not telling the public.

The Trump administration is moving to deploy custom artificial intelligence products to speed up approvals for new applications of toxic chemicals and pesticides. But it has failed to share details of these AI tools with the public as required by both federal law and the administration’s own regulations, a Monitor review has found.The new AI products for chemicals and pesticides are part of a larger push by Environmental Protection Agency Administrator Lee Zeldin, who has declared “Make America the AI Capital of the World” to be a core pillar of the agency’s mission.However, they are not listed in the EPA’s most recent public registry of “current and planned uses” of AI. The online directory is mandated by Congress to ensure the public can provide feedback on proposed AI tools before they are rolled out and assess whether the government is protecting personal privacy, civil rights, and public safety as it deploys powerful new technologies. Why We Wrote This The Trump administration is designing AI tools to vet chemicals and pesticides for safety. Critics question whether the administration is meeting legal standards for transparency. The lack of transparency is likely unlawful, environmental advocates and legal experts tell the Monitor. They say it leaves the public in the dark about how decisions are being made about the use of potentially dangerous chemicals.“The taxpayers have a right to know,” says Tim Whitehouse, executive director of Public Employees for Environmental Responsibility, or PEER, a nonprofit watchdog organization that defends government scientists, land managers, and whistleblowers who expose environmental wrongdoing.“Otherwise,” Mr. Whitehouse says, “we’re entering a whole whole new world where government becomes further removed from the American people and captured by tools that very few people understand and that have inherent biases and flaws.”Critics say the EPA’s failure to disclose details about AI tools that vet toxic chemicals and pesticides is part of a larger pattern by the administration.In July, when the Monitor asked the Department of Health and Human Services about an AI product designed to evaluate the sponsors of unaccompanied immigrant children, evidence of it disappeared from the agency’s website.This year, Emma Lurie, an incoming computer science professor at Barnard College in New York, led a team of researchers from the University of Pennsylvania and from Haverford College that reviewed official disclosures of 5,000 federal AI products – from license-plate readers and social media scanners used by immigration agents, to customer service voicebots at the IRS and the Department of Veterans Affairs.They found that more than 100 AI products were “out of compliance” with either federal law or Trump administration regulations. Among them was an EPA tool to help prioritize inspections of facilities that generate large amounts of hazardous waste.“Faster is not always better,” Ms. Lurie says. “In a democracy, understanding how high-stakes decisions are made is essential. We want decisions made with AI tools to be trustworthy and reliable in the public interest.”In an emailed response to questions from the Monitor, the EPA press office said it would be “premature” to release information about the use of AI tools to vet chemicals and pesticides because they were not yet deployed.“The processes for chemical assessment and pesticide permitting are technical and complex,” the agency said. “While [the] EPA would like to use AI to enhance these processes in the future, it is essential that any resulting regulatory products are accurate to ensure these actions are protective. The agency is continuing to deliberate on how these tools can be best deployed to ensure their use is accurate and protective.”The EPA declined to respond to many of the Monitor’s questions, including one about the AI tool that prioritizes inspections of hazardous waste facilities, which has been deployed, and Ms. Lurie’s team found did not comply with the law.President Donald Trump’s Office of Management and Budget, which sets policy around the federal government’s use of AI, did not respond to multiple requests for comment.The agency touted the use of AI to assess chemicals and evaluate pesticide permits in its most recent budget request to Congress, as part of a $120 million effort to “leverage AI to improve EPA processes.”More than 400 workers were tasked to be part of this effort, according to the budget document.But when PEER and the Freedom of Information Group filed a request under the Freedom of Information Act in March, seeking details of the chemical assessment tool – including any prompts and data involving the herbicide dicamba and known toxic chemicals such as formaldehyde and benzene – the EPA responded that none of those records existed.The environmental groups also asked for copies of user manuals provided to EPA staff as well as predeployment monitoring records for the AI tools. The EPA said those didn’t exist either.According to a slide-deck presentation produced by an EPA scientist and viewed by the Monitor, the agency’s chemical assessment tool would use AI to review scientific literature about an existing chemical to determine its potential relevance to new uses. The AI tool would extract data and appraise the quality of that data to produce a “judgment and narrative rationale” to explain why the AI determined a new use should be approved or not.That recommendation would then be reviewed by agency scientists, who would make a final decision, according to the slide deck.In August 2025, Sean Watford, the EPA scientist who created the tool, crafted an assessment of the AI product’s effectiveness. But while an overview appears on the EPA website, the full document has not been made available to the public.The previous month, in July 2025, Mr. Watford and another EPA scientist, Daniel Schoeff, had spoken about efficiencies created by the chemical and pesticide AI tools at a forum hosted by Amazon Web Services in Washington.According to a blog post about that event, the AI tool to vet pesticides combined a number of Amazon products with Anthropic’s Claude software to produce “99 percent potential reduction” in processing time.“Tasks that previously took four months now take seconds,” the blog post states.Assessments of potentially toxic chemicals could happen 85% faster than before. “Tasks that previously took months can now be completed in hours or days,” the blog says, claiming Amazon’s Bedrock and Anthropic’s Claude could replace 500 hours of EPA scientist staff time.Final decisions about chemical safety would still be ultimately made by humans, according to the blog post.However, PEER and the Freedom of Information Group worry that the EPA has significantly reduced the number of humans it employs, jettisoning nearly one-quarter of its workers during the first six months of 2025 alone.Last September, the EPA began shuttering its Office of Research and Development, which employed independent scientists who evaluated toxic chemicals.Dr. Watford, the computational toxicologist who designed the AI bot to assess chemicals, was among the employees who departed.“My specific concern is that when you say, ‘get to the right answer,’ who gets to determine what the right answer is?” asks Kevin Bell, of the Freedom of Information Group. With fewer government scientists, the EPA will be increasingly reliant on tests carried out by chemical manufacturers who are motivated by profit, he adds, rendering any human review of AI decisions inadequate.In its response to questions from the Monitor, the EPA press office said many scientists from the closed research office had been “re-deployed to other program offices across the agency, where they are conducting mission-critical work.”“Whether related to the future deployment of AI or other areas in which they may have experience, EPA will continue to rely on the skill sets of its career staff and scientists to fulfill its mission of protecting human health and the environment and ensure any tools used are fit for that purpose,” the EPA said. Lee Zeldin, the administrator of the Environmental Protection Agency, speaks to members of the media as President Donald Trump looks on, in the Oval Office, June 29, 2026. In a recent interview, Mr. Zeldin touted reducing the administration’s backlog of more than 14,500 pesticide reviews by almost half. The EPA told the Monitor that this was done without the use of AI.Mr. Zeldin, a former congressman from New York with little background in environmental regulation, has made AI adoption and deregulation twin goals of his tenure. Less than two months after being sworn in as administrator, Mr. Zeldin launched what he called “the greatest day of deregulation our nation has seen” with 31 actions, including reconsideration of regulations for coal, oil, and gas production as well as reconsideration of wastewater and air-quality standards.“He gets it approved, and he gets it approved fast,” said President Trump of Mr. Zeldin at a “Steel Across America” event on the White House lawn on Sept. 8, during which he also called the administrator “one of our real superstars.”“High-impact” AI toolsUnder Mr. Trump, the Office of Management and Budget issued guidance last year that requires products that have a “significant effect on rights or safety” to undergo additional scrutiny.Agencies must perform independent assessments of these “high-impact” tools prior to deployment, test their veracity and design protocols to mitigate risks. Agencies must also conduct “ongoing monitoring” and have an appeals process that includes human review.In their Freedom of Information Act request, PEER and the Freedom of Information Group asked for copies of assessments EPA performed on “high-impact” AI tools in 2025. The EPA denied that request, saying sharing “documentation on potential impacts on ... privacy, civil rights, and civil liberties” would “harm agency decision making if released.’”Research carried out by Ms. Lurie and her colleagues found that of the 43 cases of AI the EPA publicly disclosed, only two were deemed “high-impact” and therefore subject to the assessments and oversights required.Both of the “high-impact” cases, deployed in 2025, were “not in compliance” with the law, according to Ms. Lurie’s analysis, because they did not include key details about the products, including whether ongoing monitoring of its impact had been established.Last week, PEER and the Freedom of Information Group filed an appeal to the EPA’s general counsel arguing that “because chemical reviews affect public health, EPA should immediately provide answers about its AI programs.”As of Monday evening, they had not received a response from the agency.