The Department of Housing and Urban Development withheld more than 100 records about how officials tied to the Department of Government Efficiency used artificial intelligence in housing policy work, according to documents obtained by Democracy Forward and reported by WIRED. The fight matters because HUD rules govern real housing programs, and the public record still does not show what the AI systems were asked, what they returned, or how much weight officials gave those outputs.
The records list document names but not their contents. Several names suggest DOGE personnel at HUD were using AI tools around regulatory analysis and possible policy changes. One file owned by Scott Langmack was called “GPT defined Econ Analysis approach 11 10 25.docx,” according to WIRED. HUD withheld it under a rationale labeled “deliberative AI input.” Another Langmack document, “RegulatoryAnalysisPrompt.pdf,” appeared to concern prompts for regulatory analysis.
Other withheld files were described as regulatory analysis documents for HUD programs, though WIRED reported that the records do not make clear whether AI generated or shaped those materials. That distinction is the whole ballgame, and HUD’s response appears designed to keep it blurry.
DOGE’s HUD work focused on rules and contracts
WIRED previously reported that Christopher Sweet, then a third-year University of Chicago student, joined the DOGE team at HUD. Scott Langmack, who came from the property technology startup Kukun, was also part of that work, according to WIRED. HUD employees told WIRED at the time that Sweet was focused on using AI to identify agency rules that could be rescinded, or contracts that could be canceled, as part of a broader federal deregulatory push.
Some HUD staffers said employees were asked to review regulations flagged by AI for potential rescission. Others described the work as duplicative, according to WIRED. Sweet graduated from the University of Chicago in June with an economics degree. Langmack’s LinkedIn page now lists him as executive director of deregulation AI at the Office of Management and Budget, under the Executive Office of the President.
HUD cited privileges that AI does not get
FOIA law gives agencies nine broad bases to deny records, including personal privacy and trade secrets. HUD leaned heavily on Exemption 5, the provision that covers internal deliberations before an agency makes a decision. That privilege usually protects candid debate among government employees while they are shaping policy.
HUD’s FOIA office also used phrases such as “draft of AI prompt,” “deliberation of AI prompt” and “deliberative AI input” to justify withholding some records, according to WIRED. John Davisson, deputy director of enforcement at the Electronic Privacy Information Center, told WIRED that FOIA contains no AI-specific exemption. He said deliberation with a chatbot should not qualify the same way human staff discussion can, because computers are not owed candor.
One document, “DFR Template_Workflow Prompt Directory (3).pdf,” was withheld under both “deliberation of regulatory changes” and presidential communications privilege. Davisson told WIRED that this raised questions about where the prompts originated. Presidential communications privilege exists, but WIRED noted that courts generally treat it as applying to the president and close advisers.
Tori Noble, a staff attorney at the Electronic Frontier Foundation, told WIRED that prompt records are especially useful for understanding government AI use because such systems can hallucinate, reflect bias or produce wrong answers. Mark Fagan, a lecturer at Harvard Kennedy School, told WIRED that disclosure is good practice when AI is used to assess policy, although he said some prompt-based exploration may be part of an internal deliberative process.
No U.S. law currently requires federal agencies to disclose whether AI helped create rules, policies or regulations, according to WIRED. Dan McGrath, senior oversight counsel at Democracy Forward, said the government’s withholding of records about AI in policymaking raises serious transparency concerns. For now, the public can see the file names. The machinery behind the policy work remains off-screen.
This story draws on original reporting from WIRED.
