DOGE's SweetREX AI Rewrites HUD Rules; FOIA Suit Follows

There is a specific kind of governance risk that shows up when an AI system is doing consequential work but the details are locked behind FOIA fights, and Wired's reporting says that is what has been happening inside the Department of Housing and Urban Development. DOGE deployed an internal tool called SweetREX, named after the associate who built it, Christopher Sweet, described in the reporting as a young operative with no government experience who has not yet completed his undergraduate degree. The tool, which runs primarily on Google's Gemini models according to a PowerPoint pitch obtained through public records, scans federal rules and outputs a recommendation to "keep, delete, or partial delete" each one.

The scale is the interesting part. The pitch deck claimed SweetREX could compress roughly 36 hours of human legal and policy review of more than 100,000 comments into about 2.4 hours, and cut rulemaking time by at least 90 percent. At HUD, the tool reportedly reviewed more than 1,000 regulatory sections in under two weeks, part of a broader ambition to run across roughly 200,000 existing federal rules. The stated default is the part worth pausing on: sources familiar with the FOIA submissions said the operating assumption is that rules and regulations should be rescinded.

That default is a real policy stance encoded in a prompt, which is where the "government won't say how" framing becomes the point. Democracy Forward has filed a FOIA lawsuit against HUD along with the Office of Personnel Management, the General Services Administration, and the Office of Management and Budget, asking a court to force the agencies to disclose documents on how AI is being used to dismantle federal rules. Law scholars quoted in the reporting are skeptical that current models can reliably interpret dense statutory language, and at least one HUD employee said the AI misinterpreted statutes and flagged legal language as non-compliant when it was accurate.

The honest caveat is that most of what is public still traces back to a leaked PowerPoint and a handful of sources; the actual audit trail of which HUD rules the tool has recommended for deletion, and which of those a human signed off on, is exactly what the lawsuit is trying to pry loose. What the reporting does not settle is how many of those recommendations have already been enacted, or whether the same tool is quietly operating at OPM, GSA, or OMB.

The piece worth watching is disclosure. If courts side with the transparency plaintiffs, other federal AI deployments could be pulled into the same daylight, giving contractors offering auditable rule-review systems something concrete to sell against. If they do not, an AI with a rescind-by-default prompt sitting between statute and rulemaking becomes the template.