The Department of Housing and Urban Development has withheld more than 100 documents describing how the Department of Government Efficiency used AI to shape housing policy, according to a FOIA response obtained by Democracy Forward. Among HUD's stated reasons for denial: a nonexistent 'AI privilege' and a presidential communications privilege that is generally understood to cover only the president and their immediate advisers. The withheld files include documents titled 'GPT defined Econ Analysis approach 11 10 25.docx' and 'RegulatoryAnalysisPrompt.pdf.'
The FOIA response is the first concrete paper trail confirming that DOGE staff at HUD were building AI prompts to conduct regulatory analysis and flag agency rules for rescission. HUD staffers previously described being looped in to review rules that AI had already surfaced for cancellation, with some calling the process redundant. The identity of the model behind the prompts is not disclosed, though document filenames reference GPT.
The two DOGE figures at the center of the effort are now scattered. Christopher Sweet, a third-year University of Chicago student when he joined DOGE at HUD, graduated in June 2026 with a degree in economics. Scott Langmack, who came to DOGE from the property-technology startup Kukun, is now executive director of deregulation AI at the Office of Management and Budget inside the Executive Office of the President — a title that suggests the deregulation-by-AI approach piloted at HUD is being scaled across the federal government.
Key facts
- 01HUD withheld more than 100 documents about DOGE's AI use in housing policy in response to a Democracy Forward FOIA request.
- 02The agency cited 'deliberative AI input' and 'draft of AI prompt' as reasons under Exemption 5 — categories that do not exist in FOIA statute.
- 03One withheld file, 'GPT defined Econ Analysis approach 11 10 25.docx,' belonged to former DOGE staffer Scott Langmack, now at OMB.
- 04Christopher Sweet, a University of Chicago student who joined DOGE at HUD, graduated in June 2026 with an economics degree.
- 05No US law currently requires the government to disclose whether AI was used to draft or shape federal rules.
Nearly all the withheld documents were denied under Exemption 5 of FOIA, which shields the 'deliberative process' between federal employees working through policy decisions. HUD's twist was to extend that shield to AI prompts and AI outputs, labeling files as 'deliberative AI input' or 'draft of AI prompt.' Democracy Forward argues the extension has no basis in the statute.
John Davisson, deputy director of enforcement at the Electronic Privacy Information Center, said the deliberative process exemption was designed to protect candor between human federal workers debating policy — not to cover exchanges between an employee and a chatbot. He noted that in one case, HUD invoked presidential communications privilege for a document titled 'DFR Template_Workflow Prompt Directory,' which raises questions about who authored the prompts and at what level of the executive branch.
There is currently no US law requiring the federal government to disclose when AI is used to draft, assess, or modify rules and regulations. That gap matters because the tools in question are known to hallucinate, exhibit bias, and produce confidently wrong answers on technical questions. Tori Noble, a staff attorney at the Electronic Frontier Foundation, said access to the underlying prompts is the only reliable way to audit how officials are actually using AI and what harms might follow.
Mark Fagan, a lecturer at the Harvard Kennedy School, offered a partial defense of the government's position. He said that if an official queries an AI model the same way they might Google a question during policy research, that back-and-forth is arguably part of an embedded deliberative process. He also acknowledged that at this stage in AI's adoption, disclosing its use in policymaking would help build public confidence — a signal HUD is currently not sending.
“There is no AI exemption under FOIA.”— John Davisson, Deputy Director of Enforcement, Electronic Privacy Information Center
The redactions leave the public unable to evaluate whether AI-flagged rescissions were sound, whether the prompts embedded political assumptions, or whether the outputs were checked against agency data before being acted on. Democracy Forward's Dan McGrath framed the withholdings as a first-order transparency failure: existing privileges exist to protect human candor, not to obscure the influence of machine-generated recommendations on binding federal rules.
What HUD is testing here is whether the federal government can treat AI-assisted policymaking as a black box by default. If Exemption 5 stretches to cover prompts and outputs, every agency now has a template for keeping AI-driven decisions off the public record — and Langmack's move to OMB suggests the template is already spreading. The fight over these documents is really a fight over whether the next generation of federal rulemaking will be auditable at all.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
