DHS Will Use Google AI to Redact FOIA Documents

US Customs and Border Protection will use Google's AI tools to recommend which information should be blacked out of documents released under Freedom of Information Act requests, according to GitHub documents seen by The Washington Post.

The documents lay out a specific timeline. By the end of September, the Department of Homeland Security — CBP's parent agency — will apply AI to document sets that account for more than 10 percent of FOIA requests. That translates to roughly 100,000 to 140,000 requests in total.

The scale matters. FOIA is the primary legal mechanism through which journalists, researchers, and members of the public obtain federal government records. Any change to how redactions are made in that pipeline affects what information actually reaches requesters. The Justice Department's recent Epstein file release, which used redactions to black out text across millions of documents, has already put the practice of withholding information under unusual public scrutiny.

How the system will work

According to the GitHub documents, Google's AI tools will not black out text on their own. They will recommend information for redaction, and DHS employees will then review each case. The stated goal is that requesters receive what the document calls "accurate and complete" information — a phrase that sits uneasily beside the fact that the process exists to withhold portions of records.

The agency's own description frames the change as an efficiency measure. "[T]his automation is expected to significantly reduce processing times," the document states.

FOIA backlogs are a chronic problem across federal agencies, and slow processing times are one of the main ways the law's promise of transparency gets eroded in practice. If automation genuinely speeds up releases, requesters could see documents in weeks rather than after extended delays. The risk, as transparency advocates see it, runs in the opposite direction: AI recommendations could expand the volume of withheld material, and reviewers operating at machine-assisted speed may defer to algorithmic suggestions rather than scrutinize each proposed redaction.

DHS is not alone

The Washington Post's reporting found that DHS is the furthest along, but not the only agency bringing AI into the redaction process.

The Interior Department already uses Microsoft tools to redact potential attorney-client information from records. The Department of Justice is using a Veriton product called aiWARE to redact video. The Department of Health and Human Services is developing its own in-house tool, named FRED, designed to redact unknown information.

An HHS spokesperson confirmed the direction of travel. "FDA is responsibly exploring limited uses of AI to help FOIA staff process records more efficiently," the spokesperson told the Post.

Taken together, the GitHub documents and the agency statements describe a government-wide shift: AI-assisted redaction is moving from experiment to routine practice across at least four departments.

The secrecy concern

The government's framing is time savings. Critics see a different balance of risks.

"I think we have to watch the government really carefully on how they use AI to redact records, because I think it's going to be very a powerful tool for secrecy," a FOIA expert at the University of Florida told The Washington Post.

The concern is structural rather than speculative. FOIA contains numerous exemptions that permit agencies to withhold information — including material related to law enforcement, privacy, and internal deliberations. The discretion involved in applying those exemptions is exactly where AI recommendation systems would operate. An over-inclusive model, or a human reviewer who rubber-stamps its output, could shift the default from disclosure toward withholding without any formal change in policy or law.

There is also an accountability gap specific to AI systems. Traditional redactions are made by named officials exercising judgment that can be questioned, litigated, and reviewed. When a Google model recommends the redaction and a DHS employee signs off, the reasoning behind each blackout may become harder to contest — both because model behavior is opaque and because volume is high. At 100,000 to 140,000 requests, individual review of every AI suggestion is a demanding standard to maintain.

The choice of vendor adds another layer. Google's tools are commercial products, and their deployment inside the FOIA pipeline means a private company's software will help determine what the public is allowed to see about its government. The GitHub documents seen by the Post do not, per the reporting, detail the contractual terms or the safeguards attached to that deployment.

Why this story lands now

The timing is not accidental. Redactions became a subject of mainstream attention with the Justice Department's Epstein file release, in which vast quantities of text were blacked out across millions of documents. That episode demonstrated both how much the government relies on redaction as a tool and how contested its application can become.

Automating the practice at this moment raises the stakes on both sides of the ledger. Faster processing could mean more documents reaching the public sooner. Faster redaction could equally mean more information disappearing behind black bars, at a pace no human reviewer can meaningfully audit.

The end-of-September deadline gives the DHS deployment a measurable checkpoint. Whether the promised reduction in processing times arrives — and whether the volume of redacted material stays flat, grows, or shrinks once AI is in the loop — will define whether this becomes a model for transparency or a template for algorithmically scaled secrecy.