Product2 publishers3 min readPublished
Homeland Security will have Google's AI suggest redactions on up to 140,000 FOIA requests
Homeland Security plans to have Google's AI propose redactions on 100,000 to 140,000 FOIA requests by the end of September. Officers will review the suggestions, but the plan's stated payoff is faster processing, a measure that cannot show text blacked out in error.
The Product Desk · Product desk
What happened
- The tools go to Customs and Border Protection, where they will recommend what FOIA officers should redact, according to GitHub documents seen by The Washington Post.
- DHS employees will then review the cases so the requester gets what the document calls "accurate and complete" information.
- The Post found redaction AI in use or in development at other agencies, including Interior, the Justice Department and HHS.
Compiled by The Product DeskSomething wrong?How this is made
Why it matters
- decision Teams copying the suggest-and-approve setup have to decide what reviewers are scored on. The DHS description names faster processing as the payoff, so any measure of wrongly hidden text has to be added deliberately.
- exposure Over-redaction errors land on requesters, who see a black box but not what is under it, so the people with most reason to catch the model's mistakes have the least ability to.
- precedent With Interior, Justice, HHS and FDA already using or exploring such tools, a DHS rollout at six-figure volume makes AI-suggested redaction likelier to become the default FOIA workflow.
By October, if the plan holds, a FOIA officer at Customs and Border Protection will open a request and find the black boxes already drawn by Google's software [1][2]. The officer's job becomes deciding which boxes stay [3].
The pitch and the work are different things. The pitch is in the project description: "[T]his automation is expected to significantly reduce processing times" [4]. The work is a review step that exists, the document says, so the requester gets "accurate and complete" information [3]. Those two goals compete at the reviewer's desk. Each suggested box an officer stops to question costs time, and time is what the project promises to save [4].
The model can be wrong in two directions, and only one shows up. If it misses a name, the exposed text sits on the page in front of the officer. If it blacks out a sentence that should be released, the requester receives a box with no way to see what is under it. "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," a FOIA expert at the University of Florida told The Washington Post [10].
The rollout is partial. If 100,000 to 140,000 requests are more than 10 percent of the total [2], the full pool is below roughly 1 million to 1.4 million requests [11]. Most requests stay on the existing process, so DHS will have a hand-processed comparison group if it chooses to measure one. Engadget's account of the documents does not mention an accuracy target or an expected rate at which officers overturn the model.
The Post also found redaction AI in use or in development at other agencies [5]. Interior uses Microsoft tools to redact potential attorney-client information [6]. The Justice Department uses a Veriton tool called aiWARE to redact video [7], and HHS is developing its own tool, FRED [8]. "FDA is responsibly exploring limited uses of AI to help FOIA staff process records more efficiently," a spokesperson told the Post [9].
For any team putting a model's suggestions in front of a human approver, I'd sort the job on two axes. One is whether the person harmed by a wrong suggestion can see the error. The other is whether the reviewer is scored on cases closed or on corrections made. Visible errors and a correction-scored reviewer is the setup where human sign-off works the way the org chart describes. Hidden errors and a throughput-scored reviewer is where approvals are most likely to go through unexamined. Over-redaction in FOIA sits in the hidden column by its nature, and a project whose stated payoff is processing time [4] starts in the throughput row. The cheapest counterweight is a log of every suggested box an officer removes, reported as a share of all boxes the model drew.
What to watch
- Whether DHS publishes how often officers remove or add redactions relative to what the Google tools suggest.
- FOIA appeals or lawsuits challenging redactions on requests CBP processed with the AI tools.
- Whether DHS extends the tools past the 100,000 to 140,000 requests planned for the end of September.