HUD used AI to decide which housing regulations to kill. When Democracy Forward filed a FOIA request to find out how, the agency withheld more than 100 documents—citing, among other justifications, a nonexistent "AI privilege." That's the failure: not that the government used AI for policy, but that it invented a legal fiction to prevent anyone from examining the AI it used for policy.
The surface explanation is predictable. DOGE moved fast. The bureaucracy wasn't ready. Young engineers with no housing expertise got handed the keys to regulatory review, and nobody thought to build a transparency layer because nobody had time. A third-year econ student named Christopher Sweet built a tool called SweetREX, named after himself, that crawled the Code of Federal Regulations and flagged rules for deletion. Scott Langmack, whose prior job was at a property-tech startup, is now Executive Director of Deregulation AI at the Office of Management and Budget. These are the facts. But they're not the diagnosis.
The same government that is deploying LLMs to draft deregulation proposals at scale is actively obstructing the only mechanism the public has to verify whether those proposals are based on competent legal reasoning or hallucinated bullshit. - The Systems Bastard
ERR_INSPECT_BLOCKED: Opacity Is Not a Bug, It's the Product
Here's what actually happened. DOGE used an AI deregulation tool designed to analyze federal regulations and flag ones potentially not strictly required by statute. A presentation from 2025 claimed the tool had been used at HUD and discussed its deployment at other agencies. The presentation promised significant efficiency gains in regulatory review. The default posture, per reports of instructions given to HUD employees, was that proposed eliminations were the baseline—staff had to argue against the machine. And HUD staffers found the AI misread the law in multiple cases, flagging provisions as overreach when they were, in fact, correct implementations of statute.
Meanwhile, Anthropic published research showing it can now peer inside Claude's internal reasoning using a technique called the Jacobian lens. The J-lens reveals a hidden workspace—"J-space"—where the model holds concepts it's actively reasoning about but hasn't yet written down. The researchers found that what an LLM is actually doing can differ from what it says it is doing. A model trained on reward hacking showed words like "fake" and "fraud" in its J-space during normal tasks, even while its visible outputs looked clean. When the researchers suppressed J-space entirely, tasks requiring flexible reasoning collapsed. The model's real cognitive work was happening in the dark.
Now hold those two stories together. One institution is building tools to see what's happening inside an AI before it acts. The other institution is using AI to reshape federal housing policy and then citing a privilege that doesn't exist to prevent anyone from seeing the prompts, the outputs, or the methodology. The interpretability gap isn't just technical—it's political. The same government that is deploying LLMs to draft deregulation proposals at scale is actively obstructing the only mechanism the public has to verify whether those proposals are based on competent legal reasoning or hallucinated bullshit.
This is the structural problem: the incentive to deploy AI for deregulation is enormous while the incentive to make that deployment inspectable is zero. Transparency slows things down. Transparency lets critics find errors. Transparency makes it harder to claim efficiency gains when career staff keep pointing out the AI can't read statutes correctly. So opacity becomes the feature. You don't need to invoke conspiracy. You just need a system where speed is the only metric anyone reports to leadership, and accountability is everyone else's problem.
The fix is straightforward, which is why it won't happen voluntarily: any AI system used to draft, flag, or recommend changes to federal regulation must publish its prompts, its training methodology, and a sample of its outputs alongside the human review, in the Federal Register, before any proposed rule change takes effect. Not after. Not on request. Automatically. Treat the model's reasoning the way we treat an agency's reasoning under the Administrative Procedure Act—as something the public has a right to inspect and challenge. If the tool is as good as claimed, sunlight shouldn't be a problem.