DOGE's HUD AI deployment and the FOIA privilege that does not exist
HUD answered a FOIA request about DOGE's AI deployment in housing policy by citing a privilege that federal disclosure law does not recognize. The same pattern shows up inside every enterprise that adopts AI in a regulated function without a governance program. The case study and the three controls.
In the past month, the Department of Housing and Urban Development responded to a Freedom of Information Act request about its use of artificial intelligence in housing policy by invoking a privilege that does not exist in federal disclosure law. The records in question concern DOGE's deployment of AI in the agency's housing decisions. HUD cited, in part, an exemption that the courts and the relevant Office of Government Information have not recognized. The agency declined to release the documents.
The technical story, that a model was used in housing policy and the public may not be told how, is the kind of story that reads like an isolated incident in the press. It is the same story about to repeat inside every enterprise that adopts AI in a regulated function, with FOIA standing in for the regulator, the auditor, or the board's risk committee. The question HUD is now answering publicly, the question of how an AI was used and whether the answer has to be given, is a question every enterprise adopting AI in a regulated function will answer privately, often badly, unless the governance program is set up before the deployment.
The pattern: an AI in the workflow, no record of the decision
DOGE's role inside HUD, and inside the federal housing apparatus more broadly, has been to identify administrative inefficiencies. AI tools, large language models included, are well suited to a scan-many-records task. The risk is in what does not get recorded. When an AI-assisted decision produces an output that flows into a housing policy or a tenant screening rule, the audit trail that a regulator, a court, or a public-records requester would expect to see is not automatically produced by the model. It has to be produced by the program around the model.
What the public-records request, and the FOIA response, surfaces is whether the program around the model exists. If the program does not, the agency cannot answer. If the program does, the answer is available, citable, and reviewable. The HUD response, by leaning on a privilege that does not exist, suggests the program around the model is not in a state that would let the agency speak to the deployment. That is a documentation failure masquerading as a legal posture.
What the cited privilege is and why the answer is not the right shape
Federal FOIA law has nine exemptions. They are listed. None of them is a generic AI-internal-process privilege. There are exemptions for trade secrets, for law-enforcement records, for personnel files, for geological data, for inter-agency memoranda, for personal privacy, for law-enforcement techniques, for financial institutions, and for oil and gas wells. An agency that wants to withhold documents about how an AI was used has to fit the request into one of the nine. The HUD response, by all reporting on it, did not fit. It cited something that is not on the list.
For an enterprise audience, the translation is direct. When a regulator, a board, or a counterparty asks how an AI was used in a decision that affects them, the answer has to be documentable in the categories the requester recognizes. The categories that work are the same categories that have always worked: a documented model, a documented input set, a documented output, a documented reviewer, and a documented escalation path. An enterprise that has those five records can answer the question. An enterprise that has none of them has to find something to cite, and what gets cited is rarely a privilege; it is the absence of a record.
The three enterprise equivalents of the FOIA denial
Three patterns inside regulated enterprises map cleanly onto the HUD case.
First, a model that lives outside the procurement record. The tool was used in a regulated function, but the procurement file does not contain it. The tool came in through a developer install, a vendor pilot, or a contractor relationship that was never brought into the central vendor register. When the regulator asks the procurement function what AI tools are in use, the answer is incomplete. When the auditor asks the model-risk function the same question, the answer is also incomplete. The two incomplete answers do not match, and the regulator notices the gap.
Second, a decision that is AI-shaped but not AI-labeled. A recommendation flows from a model, a human reviewer endorses it, and the resulting policy decision is recorded as a human decision. The model's contribution is not in the decision artifact. When the regulator asks for the decision record, the answer is technically correct and substantively misleading. The model is invisible in the artifact that gets reviewed, and the human reviewer is, by the record, the only contributor.
Third, a record that is technically retained but not reviewable. The model produced a log, the log was kept, but the log is in a format, a schema, or a storage tier that the disclosure function cannot surface in a defensible way. The enterprise has the answer. The enterprise cannot produce the answer in a form that survives a regulator's clock. The result is the same as the first two patterns from the requester's perspective: the answer did not come, and the reason given is the absence of a usable record.
The governance program that prevents the pattern
The HUD case is a single incident. It will be litigated. The privilege citation will be tested in court, and the agency will either produce the records or be told to. The story that will repeat is the pattern, not the incident. The way to keep an enterprise version of the same story from happening is to build the governance layer before the deployment, with three commitments that close the gap.
An inventory of every AI in a regulated function, refreshed at least quarterly. Not a list of AI projects in the data science team's roadmap, but a list of every model, every AI-assisted workflow, and every vendor-supplied AI feature that touches a regulated decision. The list is the precondition for the answer to the question. If the list is not current, the answer to the question is the absence of an inventory, and an absence of an inventory is the most expensive thing a regulated enterprise can show a regulator.
A decision-record template that names the model's contribution. Any artifact that records a regulated decision has a field for the human decision-maker and a field for the AI contribution, when there is one. The template is enforced by the system of record, not by policy language. The field gets filled. If the field is empty, the decision is not recorded as AI-assisted, and the absence is visible to the disclosure function. The template makes the model's role visible in the artifact the regulator reads, which is the only place it has to be visible.
A disclosure-readiness test for every model in the inventory. A quarterly exercise where the disclosure function pulls the records for a sample of AI-assisted decisions and produces the kind of artifact a regulator, an auditor, or a public-records requester would expect. If the exercise fails, the gap is documented, the gap is owned, and the gap is closed before the next regulator arrives. The exercise is the rehearsal that turns the governance program from a document into a working system.
The procurement gap is the actual story
The HUD FOIA denial is one incident. The pattern, an AI in a regulated function with no documentable record of the deployment, will repeat in every regulated enterprise that has not built the governance layer above. The layer is not optional, and it is not something the AI vendor can build for the buyer. It is work the buyer's own governance program has to do, before the deployment, on a refresh cadence that matches the regulator's.
If you are evaluating AI tools for a regulated function inside your organization this quarter, the conversation you need to have is not which model has the best benchmark. The conversation is whether the disclosure function can produce, in the categories the regulator recognizes, a record of how the model was used, what the model contributed, and who reviewed the model's output. The enterprises that have already built that layer are the ones that get to keep using the tools. The ones that have not are the ones that will be reading about their own FOIA denial in the news.
Book a strategy consultation with our team to map the AI tools already in use inside your regulated functions against the three controls above, and to identify the highest-risk gap in your current AI governance program.