The same administration that blacked out millions of pages in the Epstein file release is now handing part of that job to AI. According to Engadget, GitHub documents reviewed by The Washington Post reveal that US Customs and Border Protection, operating under the Department of Homeland Security, is deploying Google’s AI tools to recommend what FOIA officers should redact from public records requests. By the end of September, the system is expected to cover documents accounting for more than 10 percent of DHS FOIA requests, somewhere between 100,000 and 140,000 documents in total.
The official framing is efficiency. DHS says AI recommendations will be reviewed by human employees to make sure requesters get what the agency calls “accurate and complete” information, which is a curious phrase when the whole point is deciding what to black out. The processing time reduction is real and probably significant. FOIA backlogs across federal agencies are notoriously bad, and any tool that speeds up throughput has genuine operational value. But speed and transparency are not the same thing.
What makes this worth watching closely is the scale and the pattern. The Washington Post’s reporting didn’t just surface the DHS program. It found redaction AI in use or development across multiple agencies:
- The Interior Department is using Microsoft tools to flag potential attorney-client privilege information
- The Department of Justice is using a Veriton product called aiWARE to redact video content
- Health and Human Services is building its own internal tool called FRED for redacting unspecified information
- The FDA confirmed it is “responsibly exploring limited uses of AI” to help FOIA staff process records faster
So this isn’t one pilot. It’s a quiet, distributed rollout happening across the federal government with little public debate about the standards these models are trained on, the error rates, or who reviews appeals when something gets wrongly withheld.
The concern from transparency experts is direct. A FOIA specialist at the University of Florida told The Washington Post that AI redaction tools could become “a very powerful tool for secrecy.” That’s not a fringe view. Automated systems trained to flag sensitive categories will almost certainly over-redact in ambiguous cases, because the cost of revealing too much is legally and politically higher than the cost of withholding too much. The incentives push toward more black ink, not less.
For developers and founders building in the govtech or legal AI space, this signals real procurement momentum for document processing AI at the federal level. Google, Microsoft, and Veriton are already inside. But the reputational and regulatory exposure tied to these contracts is significant. Getting redaction wrong, in either direction, will get attention fast.
