The Department of Homeland Security is expanding its AI use for FOIA processing in response to a sizable backlog and increasing number of requests, per the agency’s annual report.
In fiscal 2025, DHS started with 221,068 pending FOIA requests. After receiving more than 1 million requests and processing nearly the same amount, the agency ended the year with 245,572 pending requests.
Despite the uptick, DHS said it maintained its backlog — defined as the number of requests or administrative appeals at the end of the fiscal year that are beyond the statutory time for a response — at just 16% of total requests received.
Technology could help the agency make a bigger dent in the years to come.
“The DHS Office of Privacy continues to invest, develop, implement, and deploy new
technological advances in FOIA processing,” the agency said in its 2025 FOIA report, published Thursday. “It expects to launch additional automation tools this coming year to further improve efficiency and reduce administrative redundancies.”
FOIAs are opportunities for the public to request records from any federal agency. Federal agencies are required to disclose information requested unless it is exempt for reasons of national security and personal privacy, among others. The FOIA.gov website characterizes the process as a way to “ensure an informed citizenry” and that it is “vital to the functioning of a democratic society.”
The influx of AI in FOIA processes has raised concerns among experts who point to the potential for degraded transparency and accountability.
“AI might help expedite the process of finding documents and aid in ensuring that documents responsive to broader requests aren’t missed,” Cody Venzke, senior staff attorney with the American Civil Liberties Union’s Speech, Privacy, and Technology Project, said via email. “But FOIA is more than just a keyword search — redactions are a legal question with significant nuance (and debate), and we’ve seen over and over that AI is simply not ready for those sort of decisions.”
Per its use case inventory, DHS plans to tap the technology to “streamline the identification and redaction of sensitive content in documents related to FOIA requests.” The use case is in a pre-deployment phase and is categorized as not-high impact, meaning it will not need to meet the same risk management requirements as other use cases.
“Although both the Biden and Trump administrations established policies to ensure that government AI is tested, vetted, and monitored, we have not seen significant advancement there, prompting concern that FOIA-focused AI won’t be put through its paces,” Venzke added.
One criticism of the existing FOIA process is that agencies have a tendency to redact too much information.
“AI trained on existing responses will reflect that,” Abigail Kunkler, law fellow at the Electronic Privacy Information Center, said in an email. “Using AI to automate redactions almost guarantees that DHS will over-redact when it responds to FOIAs.”
“We need more transparency into DHS’s actions, not less,” Kunkler added.
While not listed in the itemized inventory, DHS mentions an AI platform, called RelativityOne, in its consolidated “Common Commercial AI Inventory.” The agency categorizes the platform as a knowledge retrieval system. A privacy threshold analysis approved by the DHS Privacy Office in 2024, however, identifies the document review software as a tool used to “timely respond to [FOIA] requests.” Used within Customs and Border Protection’s FOIA division, the platform has a feature that “learns” from reviewers’ decisions and then can “independently” tag documents as relevant or not.
The consolidated AI inventory provides no additional information on the type of risk management practices that are being deployed, or whether DHS considers the tool a high-impact use case.
“The benign ‘efficiencies’ language in the inventory and the substantive ‘can independently make such binary coding decisions’ language in the PTA describe the same system,” said Patrick Eddington, senior fellow in homeland security and civil liberties at the Cato Institute. “The implication is direct: the capability that would most plausibly trigger a high-impact review — a system that can independently code documents as responsive or non-responsive for FOIA — is precisely the capability the inventory description omits.”
The AI additions are increasing at a time when FOIA offices across the government have experienced major changes to their workforce.
More than 600 government information specialists, which is the job category for FOIA professionals, have departed the federal government since the beginning of the second Trump administration, according to OPM’s federal workforce data. Nearly 100 of those departures came from DHS.
“FOIA offices have been gutted, and that puts significant limits on FOIA staff’s ability to carefully check an AI’s work when they also are pressed to be ‘efficient,’” Kunkler said.
Despite concerns from think tanks and advocacy groups, the proliferation of AI in FOIA processes is likely to continue.
Last month, DHS posted a notice of intent to award Deloitte a contract to provide customized FOIA software featuring AI tools. The procurement document follows a NextGen FOIA Tech Showcase earlier this year, which included nearly 40 vendors. Of those, nearly 85% highlighted AI tools for audio masking and redactions, scans of documents to determine relevance, autopopulation of required fields and public-facing chatbots for FOIA reading rooms.
Skepticism remains about the reliability and accuracy of such solutions.
“AI simply cannot make the kinds of determinations that FOIA requires agencies to make,” Kunkler said.