I opened a FOIA packet last week and watched entire paragraphs disappear beneath black bars. You felt that hollow knot in your stomach the first time you saw names and dates erased. I want to show you what letting machines pick what you see actually looks like.
I’ve been reporting on government secrecy for years, and I’ll be blunt: this move matters. You’re about to read how agencies from Customs and Border Protection to Health and Human Services are handing parts of FOIA to algorithms—and why that should make you uneasy.
On my desk: a FOIA response returned faster—but with more black ink than sense
That faster letter came with fewer readable details.
The Washington Post reports the federal government is expanding AI to help redact FOIA requests. Customs and Border Protection is buying a Google AI tool; the Interior Department is using Microsoft systems; the Department of Justice has tapped Veritone; and HHS is building an internal system called FRED with a contractor. The General Services Administration already rolled out USAi to encourage AI adoption across agencies.
The government’s pitch is simple: process FOIAs faster. The law expects a response in 20 days, but requesters often wait months. Agencies promise employees will still review the AI’s picks—but promises and practice can drift apart when the volume of requests spikes.
Will AI make FOIA responses faster?
Yes—AI can index, scan, and suggest redactions far quicker than a human. But speed trades against nuance. Faster processing means less time for careful judgment, and that creates room for both missed disclosures and unnecessary secrecy.
At a committee hearing: lawmakers held up pages that had been blanked out
Members of Congress and victims’ advocates have been showing those redacted pages publicly.
Critics point to recent examples—like the heavily redacted Epstein files—that contained what Rep. Jamie Raskin called “tons of completely unnecessary redactions,” and in some cases redactions that paradoxically exposed victims. Those are the mistakes that turn redaction from protective shielding into harmful exposure or bureaucratic obfuscation.
FOIA allows redactions under several exemptions—national security and personal privacy among them—but human judgment matters. AI that undercorrects creates liability; AI that overcorrects turns sunlight into shadow. As David Cullier, a FOIA expert at the University of Florida, told the Post, giving algorithms this role risks handing the government a very powerful tool for secrecy.
Can AI redact too much or too little?
Yes to both. AI can miss sensitive material (undercorrect) or hide perfectly public facts (overcorrect). Humans tend to trust machines—automation bias—so errors can survive human review unless agencies build strong audit trails and independent checks.
In war rooms and operations centers: AI helped make high-stakes decisions—and sometimes got them wrong
Internal briefings and a Bloomberg investigation revealed AI failures that cost lives and nearly started wars.
Across the government, AI is already woven into heavy-duty workflows. The Department of Defense and U.S. Central Command used a Palantir system tied to a deadly strike in Iran; a CNN report described an AI chatbot falsely telling analysts a Chinese ship carried nuclear materials. Those incidents show how an automated nudge can escalate into catastrophe when humans defer too readily to a model’s output.
AI operates as a black box: it offers conclusions without a clear path to the why. Combine that with human trust in machines and you get what I call a dangerous shortcut. It’s like a librarian using a blackout marker on history, and when that marker is guided by a flawed model, the mistakes can be permanent.
How do we stop AI from hiding information?
We need transparent logs of what the AI redacted, public standards for redaction decisions, and routine external audits. Tools like Veritone, Google, and Microsoft must be subject to contract terms that require explainability and contestability. Congress and watchdogs should insist on independent testing before systems are put into production.
At the FOIA front line: requesters still file appeals and lawsuits—because the records matter
I’ve spoken with reporters, lawyers, and activists who plan appeals the moment a response lands with blocks of black.
There are practical steps: seek narrower searches, ask for redaction justifications, file administrative appeals, and prepare to litigate. Journalists can demand redaction logs and metadata; watchdogs can push for public benchmarks on AI accuracy. If agencies want to use Google’s models, Microsoft’s tools, Veritone, or HHS’s FRED, they must also accept public scrutiny of how those systems decide.
Agencies will keep arguing that AI saves time and money while humans retain final say. But the record shows that automation bias and opaque models can warp judgment. AI in FOIA is not an abstract technology question—it’s a concrete contest over what you and I get to know.
If the government hands redaction pens to machines, who will hold the machine accountable?
Are you ready to let a set of algorithms decide what you can and cannot learn about your government?