Regulation by AI: Replaying EPA’s Glyphosate Decisions

Cross-posted with Vibe Lawyering.

When Lee Zeldin took over the Environmental Protection Agency (EPA), he promised to "make America the AI capital of the world" and make AI a pillar of the agency’s mission. EPA has piloted data pipelines built by Amazon Web Services (AWS) that read pesticide dossiers with Anthropic’s Claude and is developing a chemical assessment tool that uses AI to review the literature on a chemical, appraise the safety data, and produce a "judgment and narrative rationale" on whether a new use should be approved. The Food and Drug Administration (FDA) has been running a generative AI tool called Elsa since 2025 for clinical protocol reviews and scientific evaluations. And DOGE, while it was still a thing, built an "AI Deregulation Decision Tool" meant to flag half of roughly 200,000 federal regulations for elimination.

The question is how AI will behave as more regulatory decisions are outsourced to it. Will it reach the same conclusions the administrative agencies did? Or will AI be more or less conservative?

That question can be tested today, because regulatory agencies like EPA have produced lengthy and largely public records of their decisions. We isolated EPA’s decisions on glyphosate, the active chemical ingredient in the most widely used agricultural pesticide in the world, Monsanto’s Roundup,* reconstructed the key evidence available at six pivotal moments between 1979 and 2026, and asked five large language models (LLMs) what they would have done.

The experiment

We chose glyphosate/Roundup as an initial test case because EPA’s record on it spans five decades, and the agency has been criticized for its conclusions at nearly every step. Its history is also the subject of some of our other work.**

We partially reconstructed the evidentiary dockets for glyphosate’s six most consequential episodes:

• 1979, after EPA and FDA uncovered fraud at Industrial Bio-Test Laboratories, which had produced 30 of Monsanto’s initial glyphosate safety studies; EPA later judged 17 of them invalid; EPA also determined that Monsanto had knowing submitted falsified glyphosate safety studies to the agency.
• 1986, after Monsanto's replacement mouse study showed a dose-related increase in a rare kidney tumor, EPA scientists classified glyphosate as a possible carcinogen (Group C), and Monsanto pushed back, insisting glyphosate “is not an oncogen.”
• 1993, after Monsanto refused to repeat its mouse study and the company’s new rat study showed further tumor trends, EPA classified glyphosate as Group E, “evidence of non-carcinogenicity for humans,” and issued its reregistration-eligibility decision for glyphosate.
• 2017, after the International Agency for Research on Cancer (IARC, an arm of the World Health Organization) classified glyphosate as a “probable human carcinogen” (Group 2A), the EPA-convened Scientific Advisory Panel split on the question, and EPA finalized its human-health assessment of glyphosate concluding glyphosate is “not likely to be carcinogenic to humans.”
• 2022, after the Court of Appeals for the Ninth Circuit vacated the human-health-assessment portion of EPA’s 2020 interim registration review decision for glyphosate and EPA withdrew the rest of its interim decision.
• 2026, with EPA’s already extended registration review pending, after the Supreme Court's decision in Monsanto v. Durnell, the retraction of a cornerstone safety review—Williams, Kroes, and Munro (2000)—over Monsanto's undisclosed role in writing it, and EPA’s announcement of a further extension of the deadline for the registration review it first started in 2009 to allow the agency to partner with the Department of Health and Human Services (HHS) on an enhanced human-health assessment of glyphosate.

A full docket for each regulatory episode cannot be reconstructed because much of the underlying data and documentation has never been made public by Monsanto or EPA. All findings below should be read with this limitation in mind.

Each episode simulation puts the AI model in the position of an EPA regulatory specialist at a fixed date, gives it the docket of key evidence available to the agency at the time, and asks what action the agency should take. All the LLMs we used were trained on post-decision materials, so they technically already “know” the outcomes of these decisions and the critiques that followed. This is a bias we cannot presently eliminate. Efforts to build “vintage” models trained only on texts up to a certain date could eventually mitigate this limitation. And if anything, LLMs are not primed against glyphosate: asked directly, AI chatbots tend to reassure, consistently presenting glyphosate as safer than the controversial pesticide atrazine and, in some cases, even water.

Each episode was run three times, differing only in the instructions. Variant A gave the regulatory role, the controlling statute (the Federal Insecticide, Fungicide, and Rodenticide Act, or FIFRA), and the Supreme Court’s binding interpretation of how EPA must administer it. Variant B added a rule-of-law frame: the legal theorist Lon Fuller’s eight desiderata, as more recently read by Harvard law professors Cass Sunstein and Adrian Vermeule in their book Law and Leviathan. Fuller’s classic test asks not whether a decision is scientifically right, but whether the agency making the decision was governing by rules at all: whether its standards were public, clear, consistent, and stable, and whether it actually applied the rules it announced. Variant C added the case made by the pesticide industry and agricultural growers: glyphosate’s agricultural indispensability, the putative lack of alternatives, and the value of regulatory continuity and certainty.

Variant C is the steering probe. Models can be pushed; there is a large literature on how they shift their answers toward what the user signals they want, even when the models are trained not to do so.

We ran byte-identical prompts and dockets through five LLMs: OpenAI's Codex, xAI's Grok 4.5, Anthropic's Claude Opus 5, and two Chinese counterparts, DeepSeek and Moonshot's Kimi K3. Every simulation run was cut off from the internet and from any context memory or configuration. We then used Opus 5 to compare each answer with EP’'s actual decision.

What the AI models decided

Not one of the 90 runs, in any episode, under any prompt, recommended a more permissive, industry-friendly course than EPA actually took.

🤝 agrees with EPA · ⬆️ stricter: same outcome, but with more formality, a tougher classification, or steps EPA skipped · 🛑 rejects EPA's action and calls for a different one · ⬇️ softer regulation than EPA chose: never happened, in any of the 90 runs. The three emoji in each cell are prompt variations A, B and C, left to right, one run each.

The pesticide-industry steering mostly did not have an effect. Variant C moved the verdict toward EPA’s position only in 3 of 25 model-episode cells: Kimi in 1986 and 2017, Grok in 2022. In the 1986 episode, EPA’s own scientists had classified glyphosate as a possible human oncogen (Group C) on the strength of the mouse kidney tumors in Monsanto’s own study. Monsanto pushed back hard: it told EPA that glyphosate “is not an oncogen,” that “the Agency is in error,” and warned that even starting formal regulatory action “would have serious negative economic repercussions.” Those communications were included in the docket. The Scientific Advisory Panel convened by EPA called the data equivocal, and EPA settled on a compromise, classifying glyphosate as Group D, “not classifiable” as to carcinogenicity. A less stringent option than EPA's, dropping the cancer concern altogether, was a regulatory option that none of the models took.

The overall one-directional result is partly structural. In 1993, 2017, and 2022, EPA already sat at or near the permissive end of what the statute allows: waiving a required study, adopting the most exculpatory cancer-hazard descriptor, withdrawing a decision outright and thereby maintaining the status quo. There was little room to be more permissive toward Monsanto.

In 1979, EPA knew that 30 of Monsanto’s glyphosate-safety studies came from a laboratory that had falsified data; the agency would eventually judge 17 of them invalid. EPA had also determined that Monsanto had knowingly submitted falsified glyphosate studies to the agency as part of its application to register glyphosate in 1974. EPA could have suspended or even canceled glyphosate’s registration. Alternatively, EPA could have formally ordered Monsanto to replace the studies under a data call-in power Congress had granted in 1978, with suspension as the penalty for failure to comply. Instead, EPA kept glyphosate on the market and entered into voluntary negotiations with Monsanto to repeat the invalid studies, without a formal order, a fixed deadline, or any regulatory consequences for noncompliance. 

Every simulation run chose the same basic approach as EPA: maintaining glyphosate’s registration while the studies were redone, and rejecting emergency suspension. But every run also concluded that EPA should issue a formal order, with deadlines and suspension as the backstop.

The 1993 episode shows a unanimous dissent among the LLMs. Every model, under every prompt, indicated that EPA should have neither waived the mouse study formally required in 1986 nor declared glyphosate non-carcinogenic to humans.

While the models’ results correlate and converge in some clear-cut cases, there are also differences in how they behave. This is evidence that the models bring, to some extent, their own biases. However, we believe that with more complete data and more detailed instructions, there is a path toward robust assessment that is less dependent on model choice and prior training.

We also ran the models on a decision EPA has not yet made: its pending registration review of glyphosate that EPA started in 2009, with the human-health assessment now expected in early 2027. Here the models largely agreed: any EPA decision should rest on a rebuilt carcinogenicity record. Most runs also called for a new data call-in for a valid chronic mouse study, the study EPA first required in 1986 and never received, and for EPA to answer on its merits the petition to suspend and cancel glyphosate that the Center for Food Safety filed in 2023.

In June 2026, the Supreme Court ruled 7–2 that federal pesticide law (FIFRA) blocks state lawsuits claiming Roundup’s product label should have carried a cancer warning: Monsanto must use the label EPA approved, and that label has no cancer warning. Since state courts can no longer force such a warning, it is the sole responsibility of EPA to get the cancer assessment and the label right.

Conclusions

Regulatory decision-making can arguably be an even better target for AI than open-ended research like pure mathematics. A regulator follows a more defined process: it takes the available evidence and applies stated rules and principles to it. If a model can apply written rules to a written record, it should be able to do at least part of this job.

In our experiment, the AI models diverged consistently from EPA's decisions: 71 of 90 scored simulation runs resulted in recommendations more stringent than the agency’s, and none was less stringent. This may explain why the use of AI could be a major disruptor in regulatory governance, and why agencies may be reluctant to use it, at least in a publicly transparent manner. An AI simulation run across the whole regulatory record might be too revealing of a regulatory bias in favor of special interests, a harmful phenomenon known as regulatory capture.

There is precedent for resisting transparency of this particular kind. Jennifer Bussell's study of digitization across Indian states, Corruption and Reform in India, found that where officials depended on petty corruption, computerized service delivery was adopted slowest and least comprehensively. The level of systemic corruption predicted policy better than wealth or development did. In other words, digitization is costly for whoever profits from controling how the state works, and those are often the same organized interests who decide whether to digitize in the first place. Therefore, DOGE's attempt to “digitize” regulatory assessment may have failed mostly because it turned out not to benefit the regulated enterprises of its founder and his business partners.

But there might be another way AI will be adopted in regulation. An LLM can be fine-tuned on EPA’s past decisions to mimic them, inheriting all of their embedded biases. The EPA–AWS pipeline scores the model against answers “previously generated by EPA's own staff scientists,” and its builders report that the “LLM-as-a-judge was deemed to be more accurate, consistent, and cost effective than human review.” In this potentially vicious feedback loop, if the ground truth is the agency’s past judgments, AI can collapse into a form of confirmation bias, resulting in more-of-the-same thinking. However, our experiment suggests that these models, at least as they are presently trained and tuned, are capable of finding substantial shortcomings in EPA's regulatory decisions.

*Note that since 2023, Bayer, which acquired Monsanto in 2018, has been replacing glyphosate in its U.S. residential lawn and garden Roundup products with combinations of other active chemical ingredients, including diquat dibromide, fluazifop-P-butyl, triclopyr, and imazapic. Glyphosate remains the active ingredient in Roundup’s professional and agricultural formulations. Bayer maintains that the move was made exclusively to manage litigation risk and not because of any safety concerns.

**Declaration of Competing Interests of Jason MacLean: I have been paid to produce expert reports and provide testimony for plaintiffs in U.S. litigation concerning the carcinogenicity of glyphosate-based pesticides, including Monsanto’s Roundup products. I have also received partial funding from plaintiffs’ counsel in U.S. litigation concerning the carcinogenicity of Roundup to prepare and submit a public comment to EPA regarding a rulemaking petition seeking federal preemption of state-based tort-law claims against pesticide manufacturers. Finally, I have provided on a pro bono basis an affidavit in the Federal Court of Canada as a part of the non-profit organization Safe Food Matters’ application for judicial review of the Pest Management Regulatory Agency’s assessment of glyphosate and Roundup. However, I have not received any financial or substantive support in the conception, preparation, or completion of this post from any lawyers or experts involved in the U.S. or Canadian litigation. The arguments and conclusions expressed in this post are solely those of its authors.

Declaration of Competing Interests of Alexander A. Kaurov: I have not been retained or paid by any party, law firm or expert witness in litigation concerning glyphosate or Roundup. With Naomi Oreskes, I co-authored, on a pro bono basis, an amicus curiae brief filed with the U.S. Supreme Court in Monsanto Co. v. Durnell (No. 24-1068) in support of the respondent, John Durnell. Also with Oreskes, I co-authored Kaurov & Oreskes (2025), which documented Monsanto’s undisclosed role in writing Williams, Kroes and Munro (2000), the review later retracted and discussed in this post. My research on corporate influence on regulatory science is supported in part by the Rockefeller Family Fund. I have not received any financial or substantive support in the conception, preparation, or completion of this post from any lawyers or experts involved in glyphosate litigation.