DOGE’s AI Plan to Kill 100,000 Rules Has a Bark Bigger Than Bite

Artificial intelligence models aren’t yet sophisticated enough to implement a Trump administration plan to repeal as many as 100,000 government regulations, according to current and former government officials and scholars who study the intersection of law and AI.

The AI-deregulation plan also faces legal hurdles. When agencies repeal regulations, they must show their actions are well-reasoned, and AI models often fail to disclose or explain decision-making processes, leaving conclusions vulnerable to legal challenges.

“It’s naive both on the technology side and the side of understanding regulation,” said Cary Coglianese, a University of Pennsylvania law professor who studies regulations and technology.

While machine learning could certainly help regulators with basic research—even spot patterns that humans miss—AI models can’t yet conduct the vast repeal of regulations that DOGE envisions, Coglianese said.

The Department of Government Efficiency in July created a plan for the rapid repeal of regulations using AI and is working with agency leaders to meet the lofty goal of repealing half of all regulations on the books by Jan. 20, 2026, the anniversary of Trump’s inauguration, according to internal documents first reported by the Washington Post.

But many government endeavors are so complicated that AI may not be able to replicate the bulk of the work.

One example is a greenhouse gas rule that the Environmental Protection Agency announced last month it would rescind. The vast, complicated history of that rule makes it unlikely that AI could write a replacement regulation without misconstruing key details, said Susan Dudley, administrator of the US Office of Information and Regulatory Affairs in the George W. Bush administration.

First, the agency must repeal the legal basis for treating greenhouse gases as a threat to public health, based on extensive peer-reviewed research. That basis has been challenged in dozens of lawsuits and upheld by a federal appeals court.

AI models “may be able to succeed at public notices in the Federal Register, but they’re likely not going to stick without a more thorough effort,” Dudley said.

DOGE allies say that the medium shouldn’t matter.

Former DOGE General Counsel James Burnham said courts should judge rules on the quality of the final product and nothing else.

“It feels like saying if you do it on a computer instead of by hand, somehow that is bad,” Burnham said in an interview. “It doesn’t make any difference.”

Scott Kupor, director of the Office of Personnel Management, said Monday that his office wants to use AI to summarize comments on proposed rules and act as customer support on health-care and retiree benefits for federal workers. The goal is to boost productivity by at least 15%, he said—not eliminate humans from the process.

“We literally have people who are manually reading 40,000 comments and writing responses,” Kupor said. “Somebody wrote a response that is 70 pages long to one of these comments. Now, that’s just inefficient.”

Repealing by Automation

In a PowerPoint, dated July 1, DOGE said it would eliminate 50% of all regulations, an effort made possible by a new AI tool that DOGE says is 15 times faster than human-only review.

The presentation claims that DOGE used an AI tool in efforts to deregulate the Consumer Financial Protection Bureau and the US Department of Housing and Urban Development.

Some states are exploring similar ideas. Virginia Gov. Glenn Youngkin (R) has begun a sweeping review of the state’s administrative code using AI.

The tool can scan thousands of pages of regulatory text and identify where the language differs from the law, or where it’s more cumbersome than in other states, said Reeve Bull, director of Virginia’s Office of Regulatory Management.

Regulators don’t use it to write text or make significant decisions, he added. In the DOGE plan, lawyers edit the AI-generated text.

“There’s always a human deciding at the end of the day,” Bull said.

The Administrative Procedure Act requires the government to show that its actions are well-reasoned, which could be hard to prove with an abundance of AI involvement, Coglianese said, in part because deep-learning tools use where users can’t see the decision-making process.

However, Burnham said AI may be especially suited for deregulation because the process isn’t as complicated as writing new rules from scratch. Furthermore, it can help manage huge troves of comments from outside groups, a growing number of which are also generated by AI.

‘Starting Point’

In a statement, a senior administration official said AI “can dramatically increase the efficiency of this process, but it serves merely as a starting point for human decision-makers.”

“AI can help regulators determine what they want to cut, and it can assist lawyers in analyzing statutory text and assessing potential arguments both for and against rescinding regulations,” the official continued. “It does not supplant human judgment.”

While most scholars agree that AI may be helpful to shape regulations, they say the DOGE plan tests the boundary of what’s acceptable.

“It takes the most extreme version, the most risky version of AI that could be applied to regulations,” said Bridget Dooling, an administrative law professor at Ohio State University, who previously worked for the George W. Bush administration.

Consumer AI tools have been plagued with errors, presenting information that is wrong or fabricated. AI “hallucinations” can arise from flawed or misinterpreted data sets.

Much like Virginia, DOGE says its tool will identify which regulations are required by law and which aren’t—the starting point for mass deregulation.