The US Department of Housing and Urban Development used an AI tool, built inside government by a DOGE-affiliated team, to scan thousands of housing regulations and recommend which ones to cut. The tool is called SweetREX. It was developed by Christopher Sweet, who at the time was a third-year economics student at the University of Chicago, and Scott Langmack, who joined from a private property data company called Kukun. Langmack now holds a senior deregulation AI role inside the White House budget office.
The tool ran on Google's Gemini AI models and was designed to go through all of HUD's rules and return one of three verdicts: keep, delete, or partially delete. Attorneys were supposed to review those verdicts, and agency staff would have the final say. But documents obtained through a public records request tell a different story about the starting point. Internal instructions told HUD employees that regulatory elimination was "the default case" and asked them to flag disagreements, not the other way around.
At HUD alone, the AI reviewed more than 1,000 regulatory sections in under two weeks. A HUD employee told reporters that the AI misread statutes and incorrectly flagged legally sound rules as non-compliant. Another employee described going through AI-flagged regulations to justify why they should stay as "painful." An expert who previously worked at the Department of Labor said DOGE's assumptions about how long regulation reviews take were "wildly incorrect," noting that real review processes take weeks, not the 2.4 hours the tool's internal projections claimed for 100,000 public comments.
The same tool was being demonstrated to staff at the Environmental Protection Agency, the State Department, and the Federal Deposit Insurance Corporation, meaning its reach goes well beyond housing.
When a legal advocacy group called Democracy Forward filed a public records request to understand how this AI was being used in policy decisions, more than 100 documents were withheld. The government cited a legal exemption for "deliberative AI input," a category that legal experts say simply does not exist under US law. The deliberative process exemption, which is a real and legitimate legal protection, is meant to let government employees debate policy candidly without fear. Legal experts say it was never designed to cover conversations between a person and an AI chatbot, because an AI has no legal right to candor.
One withheld document was titled "GPT defined Econ Analysis approach," belonging to Langmack. Another was called "RegulatoryAnalysisPrompt.pdf." A third document was blocked partly on presidential communications privilege, which normally applies only to the president and their closest advisers, raising questions about where the AI prompts originally came from.
There is no US federal law requiring the government to disclose when AI has been used in writing or changing regulations. That gap is the core issue here. Courts reviewing whether an agency properly followed the law when changing a rule require agencies to show "reasoned" decision-making. Legal scholars have already noted that a process driven by an AI tool, with minimal human review, may not meet that standard. Several regulatory rollbacks based on AI recommendations could be challenged in court and reversed.
For businesses operating under US federal rules, in any sector, the question is not whether AI is being used in Washington to reshape the rules that govern them. It clearly is. The question is whether the process is accurate, well-supervised, and legally defensible. Right now, none of those things can be verified, because the government will not show its work.