미 정부 AI 주택 정책 활용, 방식은 비밀에 부쳐져

Comment Loader Save Story Save this story Comment Loader Save Story Save this story Members of the so-called Department of Government Efficiency (DOGE) who were working at the Department of Housing and Urban Development (HUD) used artificial intelligence to inform policy decisions. Now, the agency appears to be denying Freedom of Information Act requests for information on the development and use of AI tools, and the way they informed policy decisions, according to documents obtained by a FOIA request by Democracy Forward, a nonprofit legal organization. Last year, WIRED reported that Christopher Sweet, who was then a third-year student at the University of Chicago, had joined the DOGE team at HUD, along with Scott Langmack , who came to DOGE from a property technology startup called Kukun. Sweet’s primary focus, according to HUD employees who spoke to WIRED at the time, was on using artificial intelligence to identify agency rules for potential rescission, or contract cancellations, as part of a similar effort across the government . At the time, HUD staffers told WIRED that employees were being looped in to give feedback on regulations that were flagged by the AI for rescission. Other employees, however, described the effort as redundant. Sweet graduated from the University of Chicago in June with a degree in economics; Langmack is now the executive director of deregulation AI at the Office of Management and Budget (OMB), under the Executive Office of the President, according to his LinkedIn. More than 100 documents that had been requested by Democracy Forward about HUD’s AI use for decisionmaking were withheld. Among the reasons HUD cited for not releasing documents were a nonexistent AI privilege and a privilege for presidential communications that is real but generally held to apply only to the president and their immediate advisers. Several of the withheld documents, whose names are shared in the FOIA but whose contents remain unknown, appear to indicate that the DOGE team at HUD was using AI tools to help make policy decisions. Sweet, Langmack, HUD, OMB, and the White House did respond to requests for comment. One document, labeled “GPT defined Econ Analysis approach 11 10 25.docx,” which belonged to Langmack, was exempted from FOIA because it was labeled as “deliberative AI input.” Another document, titled “RegulatoryAnalysisPrompt.pdf,” also belonging to Langmack, appears to indicate that the DOGE team was looking into creating prompts to conduct regulatory analysis. Several of the other documents that were withheld for being part of the deliberative process were labeled as some form of “regulatory analysis” for different HUD programs, though it is not clear if AI was used in their creation. Tori Noble, a staff attorney at the Electronic Frontier Foundation, says that the lack of transparency around how AI tools might be used in the creating or changing policy is particularly worrisome, because tools have been known to hallucinate , show bias , or just plain get things wrong . “It’s not necessarily the case that we'd always know how tools are being used,” she says. “So having access to the prompts is really the best way to be able to tell what officials are using these tools for and how harmful those uses might be.” There are currently no laws in the US that require the government to disclose if AI has been used in the creation of rules, policies, or regulation. “If AI is being used to assess policy as one of the tools in the toolkit, I think at this stage in the development and use of AI, it is good protocol to indicate that,” says Mark Fagan, a lecturer at the Harvard Kennedy School. “In part to try and build confidence in the use of AI in government.” But Fagan says that he can see a way that AI could be used in a kind of deliberative process that may not necessarily warrant disclosure. “If I am reviewing policy to decide where I want to prune or redesign or appeal or enhance policy, and I go and Google, ‘How has this been done by others?’ I don't disclose that,” says Fagan. “It is part of my embedded process.” Fagan also notes that often it may take several prompts to work through a problem or question, and so “from a technical perspective I think much of it is deliberative.” Since many of the documents were withheld from Democracy Forward’s request, it’s also difficult to exactly ascertain how AI was being employed. There are nine different reasons the government can deny a FOIA request, including, for instance, if a document would reveal personal information or trade secrets. Nearly all of the documents from HUD are labeled as withheld due to Exemption 5, which allows agencies to withhold information under what’s known as “deliberative process privilege.” This protects material where federal workers may be workshopping or giving feedback on new or changing policy and regulations. “If there is a deliberative process, occurring between human beings that work for the agency, where the material is both pre-decisional and leading up to some sort of decision, that might be a justifiable assertion of deliberative process privilege there,” says John Davisson, deputy director of enforcement at the Electronic Privacy Information Center. Davisson says this is meant to encourage “candor” so that government employees can debate policies and give their honest feedback without fear. In many cases, the reason for the exemption is listed as “deliberation of process” or sometimes “attorney-client privilege.” But in a handful of cases, HUD’s FOIA office denied the request for certain documents, citing “draft of AI prompt” and “deliberative AI input” as reasons. Documents labeled “Prompt.pdf” and “PROMPT+AB2(alr)+ab.dox” were exempted because they were labeled “deliberation of AI prompt.” Similar exemptions occurred for a number of other documents. “There is no AI exemption under FOIA,” says Davisson, noting that any deliberation done between a person and an AI chatbot should not qualify for the exemption, because “AI systems, computers are not entitled to candor.” Another document labeled “DFR Template_Workflow Prompt Directory (3).pdf” was exempted because it involved “deliberation of regulatory changes” and “presidential communications privilege.” This reason, according to Davisson, raises “questions about where the prompts may have come from.” “It's concerning that the government is shielding records about how it uses AI in policymaking. If the government is going to use AI in formulating policy that affects us all, the public has a right to understand its impact,” says Dan McGrath, senior oversight counsel at Democracy Forward. “Withholding AI input into the policy process as deliberative and privileged raises serious questions. Existing privileges are meant to ensure that public officials share their candid views, not to shield the impact of AI's impact on our government.”