Government AI shields policy decisions from public view
The year is 2025, and the U.S. Department of Housing and Urban Development (HUD) has been using artificial intelligence to decide which federal regulations to keep, cut, or rewrite. This is not a future scenario from a sci-fi novel. It is happening now, documented in Freedom of Information Act (FOIA) requests filed by Democracy Forward, a nonprofit legal organization. The documents obtained reveal a troubling pattern: the government is not just using AI to inform policy — it is actively shielding how it does so from public scrutiny. This is where AI makes us superfluous, not by replacing our jobs, but by replacing our right to know how decisions that affect millions of lives are made
The AI That Decides What Rules Matter
At the center of this story is the so-called Department of Government Efficiency (DOGE), a team embedded within HUD. Two key figures emerged: Christopher Sweet, then a third-year economics student at the University of Chicago, and Scott Langmack, a former employee of a property technology startup called Kukun. According to reporting by WIRED, Sweet’s primary assignment was to use artificial intelligence to identify agency rules for potential rescission or contract cancellation. This was part of a broader government-wide effort to streamline regulations, but the method was anything but transparent.
HUD employees told WIRED that they were brought in to give feedback on regulations flagged by the AI for elimination. Some called the effort redundant. Others were simply looped in after the machine had already made its recommendations. The process was not collaborative; it was hierarchical, with the AI at the top. The human experts were reduced to quality control, checking the work of an algorithm they did not design and could not fully understand. This is the first layer of being made superfluous: not by being fired, but by being demoted from decision-maker to validator
The Privilege That Doesn’t Exist
When Democracy
Forward filed its FOIA request for documents related to HUD’s AI use, the agency withheld more than 100 records. Among the reasons cited were a “nonexistent AI privilege” and a “presidential communications privilege” that is real but generally applies only to the president and their immediate advisers. The documents themselves were labeled with phrases like “deliberative AI input” and “draft of AI prompt.” One document, “GPT defined Econ Analysis approach 11 10 25.docx,” belonged to Langmack and was exempted under this novel category. Another, “RegulatoryAnalysisPrompt.pdf,” suggested the team was creating prompts specifically to conduct regulatory analysis.
The deliberative process privilege, which protects internal government discussions from public view, was designed for human beings. It exists to encourage candor among federal workers who need to debate policies without fear of immediate backlash. But AI systems, as Davisson noted, “are not entitled to candor.” They do not have feelings, careers, or reputations to protect. Using a privilege meant for human deliberation to shield machine-generated input is a category error — and a convenient one for those who prefer to operate in the dark.
The Prompts That Shape Policy
The documents that were released, even in redacted form, reveal a pattern. Many were labeled as some version of “regulatory analysis” for different HUD programs. It is not always clear whether AI was used in their creation, but the pattern is suggestive. When the government denies a FOIA request citing “deliberation of AI prompt” or “deliberative AI input,” it is essentially admitting that AI played a role in the deliberative process — and then claiming that role is protected from disclosure.
Tori Noble, a staff attorney at the Electronic Frontier Foundation, explained why this matters. AI tools are known to hallucinate, show bias, or simply get things wrong. Without access to the prompts, the public has no way to verify whether the AI was fed accurate data, whether its outputs were checked for errors, or whether it was used to justify predetermined outcomes. “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,” she said. The prompts are the Rosetta Stone of AI-driven policy. Without them, the entire process remains opaque.
The Legal Void
There are currently no laws in the United States that require the government to disclose if AI has been used in the creation of rules, policies, or regulations. This legal vacuum creates a perverse incentive: agencies can use AI to shape policy, then hide that fact behind existing exemptions designed for human deliberation. That casual search, he argued, is part of an embedded process that does not require public reporting.
But there is a critical difference between a human Googling for context and an AI generating recommendations that are then used to justify regulatory changes. The AI does not have a brain to question, a background to evaluate, or a bias to disclose. When a human Googles, the search history is their own. When an AI generates an analysis, the underlying data, training, and prompt engineering are proprietary or hidden. The comparison collapses under scrutiny. The government is not just using a tool; it is outsourcing judgment to a system that no one inside the agency fully understands.
The Historical Precedent We Are Ignoring
In the 1970s, the rise of computerized data collection raised similar concerns. The Privacy Act of 1974 was a direct response to fears that federal databases could be used to track, profile, or discriminate against citizens. That law required agencies to disclose what information they collected, how it was used, and who had access to it. It was a hard-won compromise between efficiency and accountability.
Today, we are repeating the same pattern with AI, but without the legislative safeguards. The Privacy Act was passed because Congress recognized that new tools required new rules. The absence of such rules for AI means that agencies can adopt the technology without any corresponding obligation to explain how it works or what it decides. The historical lesson is clear: technological innovation without legal guardrails leads to abuse. The current situation at HUD is a textbook example of this dynamic playing out in real time.
The Mechanism That Misses the Problem
The government’s reliance on FOIA exemptions to hide AI use reveals a deeper structural flaw. The deliberative process privilege was designed to protect the free exchange of ideas among human experts. It assumes that the back-and-forth of debate, the drafting of memos, the editing of proposals — all of this requires confidentiality to function properly. The privilege is a shield for human creativity and candor.
But AI does not need candor. It does not need to be protected from embarrassment or political pressure. When an AI generates a prompt, it is not deliberating in any meaningful sense. It is executing a set of instructions. The government is using a privilege meant for human psychology to protect a process that is purely mechanical. This is not a loophole; it is a category mistake that undermines the entire purpose of FOIA. The mechanism of transparency is being applied to the wrong object, and the result is secrecy by design.
The Real Cost of Secrecy
Dan McGrath, senior oversight counsel at Democracy Forward, summarized the stakes: “If the government is going to use AI in formulating policy that affects us all, the public has a right to understand its impact.” The documents withheld from Democracy Forward’s request include not just prompts but also analyses that may have been produced by AI. Without knowing what the AI was asked, what data it was given, or how its outputs were interpreted, the public cannot evaluate whether the resulting policies are sound, fair, or lawful.
The cost of this secrecy is not abstract. Housing policy determines who gets loans, who gets evicted, who gets subsidies, and who gets left behind. If an AI is used to identify regulations for rescission, and those regulations were designed to prevent discrimination or ensure safety, the consequences could be severe. A biased or hallucinating AI could recommend cutting protections that took decades to build. And the public would never know why.
The Technical Solution That Is Socially Unresolved
From a purely technical standpoint, the problem is solvable. Agencies could be required to log all AI interactions used in policy development, store those logs in a searchable database, and make them available through FOIA or even proactively. The technology for audit trails exists. The infrastructure for transparency is already in place. The legal framework for FOIA has been tested for decades. The technical solution is straightforward: treat AI prompts and outputs as government records subject to the same disclosure rules as any other document.
But the social resolution has not followed. The political will to mandate such transparency is absent. The incentives for agencies to hide their AI use are strong: secrecy allows them to avoid scrutiny, deflect criticism, and proceed with policies that might not survive public debate. The technical fix is ready; the social fix is not. This is the crux of the problem: we have the tools to make AI transparent, but we lack the collective commitment to use them. The technology is ahead of the governance, and until the governance catches up, the black box will remain closed.
The Precedent That Should Guide Us
The Privacy Act of 1974 offers a model. It was passed in response to a specific crisis of confidence in government data practices. It required agencies to publish notices about their data collection, to allow individuals to access their own records, and to correct errors. It was not perfect, but it established a baseline of accountability that had not existed before. The same approach is needed for AI. A law requiring agencies to disclose when AI is used in policy decisions, to log the prompts and outputs, and to subject those logs to FOIA review would not solve every problem, but it would close the current loophole.
The alternative is a world where government decisions are made by algorithms that no one can question. That world is not hypothetical. It is being built right now, one withheld document at a time. The question is whether the public will demand to see inside the black box before the box becomes the only way decisions are made. The technical solution is waiting. The social resolution is not. And that is where AI makes us superfluous: not by replacing us, but by rendering our scrutiny irrelevant