CBP to Use Google AI to Suggest FOIA Redactions, With Humans Still Reviewing
U.S. Customs and Border Protection will use Google's AI tools to recommend information for FOIA officers to redact.
CBP, part of the Department of Homeland Security, is applying the system to Freedom of Information Act responses. That law lets the public request federal records, and redaction means blacking out sensitive parts before release. DHS planned by the end of September to use AI for documents making up over 10 percent of FOIA requests, totaling around 100,000 to 140,000, according to Engadget. DHS separately described the effort as a way to streamline the identification and redaction of sensitive content when responding to FOIA requests, as reported by The Washington Post.
The workflow keeps a person in the decision chain. DHS says employees will review AI redaction recommendations to ensure requesters receive "accurate and complete" information. DHS says the automation is expected to significantly reduce FOIA processing times.
CBP is not operating in isolation. The Interior Department uses Microsoft tools to redact potential attorney-client information, the kind of private legal advice that can be legally withheld. The Justice Department uses Veritone's aiWARE tool to redact video. The Department of Health and Human Services is developing its own tool called FRED to redact information. An FDA spokesperson said FDA is responsibly exploring limited uses of AI to help FOIA staff process records more efficiently.
A longer FOIA technology pipeline
The current deployments follow several years of administrative work on FOIA processing technology inside the Justice Department. The mission of the DOJ Office of Information Policy is to encourage and oversee agency compliance with the Freedom of Information Act. The office is responsible for developing government-wide policy guidance on all aspects of FOIA administration. It provides legal counsel and training to agency personnel on FOIA. It publishes the United States Department of Justice Guide to the Freedom of Information Act. Kathryn Blankenberg is the Director of the office.
That office hosted a NexGen FOIA Tech Showcase covering Artificial Intelligence tools that may assist with FOIA case processing, according to DOJ. It later sought vendors for NexGen FOIA Technology Showcase 2.0 to present redaction tools including automatic redaction of similar forms and record types including paper, digital content, and videos, according to DOJ. The Chief FOIA Officers Council Technology Committee published working group charters including Artificial Intelligence Video Review and Redaction.
The policy context extends beyond FOIA shops. The White House plans to hire 100 AI professionals to promote the safe use of AI, according to Reuters. Open government and civil rights advocates warn that using AI to answer Freedom of Information Act requests may create new problems, according to NBC News.
The broader context here is familiar to anyone who has run document-sorting software in daily work. Redaction looks like a contained prediction task, spot the passages that fall under a legal exemption and flag them, but the mistakes cut in different directions. Over-redaction holds back material the public should see and weakens trust. Under-redaction lets out protected material and can create legal exposure. Human review can catch both kinds of error, but only if reviewers have the time, tools, and incentive to overrule the system instead of just accepting its suggestions.
Looking at daily practice, the questions that will matter are operational. How often do officers accept, change, or reject the recommendations. How disagreements are logged and used to check performance. How accuracy varies across document types, from structured forms to free-text email threads to video. How decisions about privilege are handled when the context runs across several records. Those numbers will show whether processing gets faster without quietly changing how much is released.
In my view, the use of different vendors is worth flagging. Google models at CBP, Microsoft tooling at Interior, Veritone for video at Justice, and a custom FRED build at HHS suggest agencies are matching tools to the kind of records they hold and the systems they already use, rather than settling on one federal redaction model. That is a practical fit for varied records, though it makes government-wide auditing and comparison harder. The DOJ showcase and working group structure may offer a place to share ways of testing, even when the underlying systems differ.
For FOIA requesters, little changes on the surface. Requests go in through the same channels. Responses arrive with redactions marked under statutory exemptions, the specific legal rules for withholding. The difference sits upstream, in how candidate redactions are found for review. If DHS meets its September scope and keeps reviewer oversight in place, the near-term test will be simple. Faster queues, with release quality held constant.
