Customs and Border Protection is reportedly acquiring Google AI and machine-learning tools to recommend which portions of records should be redacted before public release. According to reporting by The Washington Post and Engadget, DHS expects the technology to apply to more than 10 percent of its FOIA requests, covering an estimated 100,000 to 140,000 submissions. At that scale, this is not a peripheral experiment.
The Department of the Interior is reportedly using Microsoft technology to identify potential attorney-client material for redaction. The Department of Justice is reportedly using Veritone software to assist with redacting video, with a FOIA employee required to review processed content before any final determination. At the Department of Health and Human Services, an AI tool reportedly called FRED is in development with an unnamed contractor, though the reporting does not specify which categories of information FRED will flag.
Human employees are still expected to review proposed redactions before any final determination is made, and these systems carry no formal legal authority to make binding redaction decisions. Not every agency describes its system in identical terms: DOJ’s documented approach involves automated processing followed by required human review, rather than a pure recommendation model. What the reporting does not establish is how agencies plan to measure whether that human review is substantive, consistent, or auditable.
FOIA operates on a presumption of disclosure: agencies must release requested records unless one of nine statutory exemptions applies. Those exemptions cover classified national-defense information, personal privacy, confidential commercial data, law-enforcement material, and several other categories. The law requires agencies to respond within 20 working days, a deadline that can be extended under defined unusual circumstances.
“A very powerful tool for secrecy.” That warning comes from David Cuillier, a FOIA expert at the University of Florida, as reported by The Washington Post, describing what AI-assisted redaction could become if agencies rely on it to reinforce overly broad withholding decisions.
The Gap Between Speed and Accountability
Faster processing and reliable disclosure are not the same thing, and the difference matters to anyone who depends on FOIA.
Consider the two directions in which AI recommendations can fail. Undercorrection means releasing information that should legally remain protected. Overcorrection means redacting information that should be public, leaving requesters with responses that tell them very little.
Automation bias sharpens that concern. When an AI system flags material for redaction and presents its output with apparent technical precision, reviewers tend to accept those recommendations at elevated rates. That acceptance rate can exceed what the same suggestion from a human colleague would earn. This pattern has been observed across high-stakes decision contexts, from hiring to medical diagnosis, and there is no obvious reason government document review would be immune.
A DHS spokesperson stated that the goal is to “ensure that the requester is getting the accurate and complete information requested,” as reported by PCMag. That framing describes the intended outcome. It does not describe what happens when the AI gets it wrong, or how a requester would know.
The available reporting does not establish that any named agency publishes error rates or maintains auditable logs of accepted or rejected AI recommendations. Nor does any agency, based on the reporting, appear to disclose to requesters that AI played a role in their response. That gap matters, because without baseline transparency, the assurance that humans remain in the loop functions more as a procedural description than a genuine accountability mechanism.
FOIA redaction is one piece of a broader federal movement toward AI-assisted administration, spanning reported uses in regulatory work, public health, transportation, and other agency functions. That context does not make the FOIA application less consequential. If anything, the breadth of adoption makes the absence of uniform standards across agencies more worth examining.
What Meaningful Oversight Would Actually Require
Speed and accuracy can coexist, but only if agencies build the structures that make accountability possible.
Agencies would need to publish error rates for their AI systems and maintain logs of recommendations that were accepted or rejected, making that data available to requesters and oversight bodies. Review procedures would need to be substantive enough that a human examiner is genuinely evaluating each recommendation rather than confirming outputs under deadline pressure. Requesters who believe a redaction was influenced by an AI error would need a meaningful appeal pathway. The existing administrative review process was designed before these tools existed and was not built with them in mind. The statutory exemptions have not changed. What has changed is what, or who, is doing the initial work of applying them, and whether the public can see how that work is being done.
