The Pulse
DHS Plans Google AI Redaction Recommendations for FOIA Requests
Customs and Border Protection plans to use Google AI and machine-learning tools to recommend redactions in some Freedom of Information Act responses, according to a Washington Post report. DHS says employees will review cases, while questio
AI.info Team · Written with AI ·
“DHS employees will review the case and ensure that the requester is getting the accurate and complete information requested.”
A DHS spokesperson, quoted by The Washington Post
The Department of Homeland Security plans to use Google AI and machine-learning tools to recommend what information should be withheld from some public-records responses. The Washington Post reported the plan on September 26, citing a federal AI inventory and comments from a DHS spokesperson. The Office of Management and Budget’s repository is a compilation of agency inventories, not DHS’s original inventory. The Post identified Customs and Border Protection, a DHS component, as the agency applying the tools to Freedom of Information Act processing.
The recommendations are not described as automatic final redactions: DHS says employees will review cases before records go to requesters. But the plan puts software inside a decision process that determines what the public can learn from government files—and whether a proposed withholding fits an exemption under federal law.
Up to 140,000 requests in scope
DHS told the Post that by the end of September it planned to use AI for certain categories of requests representing more than 10 percent of its FOIA submissions—approximately 100,000 to 140,000 requests. The estimate refers to the volume of submissions in those categories, not a confirmed count of documents already processed by the system. The Post reported that the inventory entry says automation is expected to reduce processing times.
The Post’s account describes Google tools as recommending redactions to FOIA officers. It does not establish that the system makes final withholding decisions, and the report does not specify a Google model, the records that will be fed into it, or how the agency will measure its recommendations. Those details matter because FOIA exemptions cover different kinds of protected information, and a recommendation to withhold still requires an agency decision.
Human review meets a large redaction workload
The scale of DHS’s FOIA work is substantial. In fiscal year 2025, the department processed 994,992 requests, according to the latest figures cited by the Post. It released records without redactions in 2.3 percent of those cases; about 40 percent included records with at least some redactions.
Those figures do not show that every redaction is improper: FOIA permits withholding information under nine statutory exemptions. But they make the quality of the review consequential. If AI recommendations help staff move through cases faster, the practical question is whether that speed comes with consistent application of the exemptions and careful review of what can be released.
Evidence on automated redactions is limited
David Cuillier, director of the Freedom of Information Project at the University of Florida’s Brechner Center for the Advancement of the First Amendment, warned the Post:
“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 a very powerful tool for secrecy.”
David Cuillier, director of the Freedom of Information Project at the University of Florida’s Brechner Center for the Advancement of the First Amendment, quoted by The Washington Post
Jason R. Baron, the former director of litigation at the National Archives and Records Administration, said agencies remain “in the early days” of using sophisticated AI in FOIA workflows, while seeing potential for faster releases.
A 2022 study by Baron and two co-authors tested AI recommendations on more than 3,000 paragraphs from email records and attachments at the Clinton Presidential Library. The researchers examined whether material could be withheld under the deliberative-process exemption; the AI methods achieved about 70 percent accuracy in identifying paragraphs that could legally be withheld. Baron told the Post that the system sometimes recommended withholding information improperly and sometimes recommended releasing information improperly.
The test is whether staff can explain each withholding
The National Archives’ Office of Government Information Services reported that 18.6 percent of respondents in its 2024 records-management assessment used AI or machine learning in FOIA processing. The assessment’s question focused on using those methods to organize and retrieve responsive records, not on the accuracy of AI redaction recommendations. That distinction makes the percentage a poor measure of how ready agencies are to automate decisions about exemptions.
FOIA’s foreseeable-harm standard says agencies may withhold information only when disclosure would foreseeably harm an interest protected by an exemption. Human review offers a checkpoint, but its value depends on whether staff can evaluate each suggestion rather than simply approve an automated result. For requesters, the output will be visible in the released record: either fewer delays with necessary redactions, or black bars that arrived faster.