The results are in: According to their own reports, some federal agencies are already testing AI in their FOIA offices. Many more are already using machine learning functions baked into current document search and discovery tools or plan to make use of AI in the future.
Volunteers responded to MuckRock’s call for help and reviewed Chief FOIA Officer Reports submitted to the Attorney General in 2023 and 2024, gleaning how different agencies are using or plan to use artificial intelligence in their FOIA process. Each federal department or agency receiving more than 50 requests must submit the report to the Attorney General. Starting in 2023, senior FOIA officials had to answer a question that they hadn’t the year before: “Does your agency currently use any technology to automate record processing? For example, does your agency use machine learning, predictive coding, technology assisted review or similar tools to conduct searches or make redactions?”
The answers supplied show how agencies are testing cutting-edge technologies and using more common machine learning technology that has been available for years through “e-discovery” tools, or software built to navigate large collections of documents in a legal setting.
In an age of Large Language Models (LLMs) and chatbots, these e-discovery tools with machine learning capabilities might not fit the newer definitions of AI. Adam Marshall, senior staff attorney for the Reporters Committee for Freedom of the Press, said it’s critical to differentiate between new AI tools and e-discovery tools that have long been used in law and FOIA.
“Pulling apart all of those different strands and being on the same page in terms of what we actually mean when we say things like AI is going to be really important,” Marshall said.
Exploring, acquiring and testing cutting-edge AI
Among the reviews in MuckRock’s crowdsourced data, the State Department demonstrated the most well-documented and extensive use of artificial intelligence for FOIA. In 2023, the agency released the first documents that were declassified and released with the help of decisions taught to a machine learning model.
The agency trained a model to declassify diplomatic cables, a type of communication between U.S. embassies abroad and the State Department, based on previous declassification review decisions made by humans. In their 2024 Chief FOIA Officer Report, the Department said it was already piloting more programs to make use of artificial intelligence.
The National Archives reported that it created a senior-level committee called the “FOIA and AI Executive Steering Committee,” but budget constraints have hampered the development of tools that the agency has been exploring.
But the Archives responded in more detail to an executive order asking federal agencies to inventory and share their AI use cases, describing an “AI system” that “would utilize natural language processing (NLP) to search records based on content similarity to the FOIA query” and “automatically redact sensitive information from the records.” According to the Archives’ inventory of AI use cases, this system has already been developed and is currently being tested.
Other agencies, like the Central Intelligence Agency, Department of the Treasury, Department of Justice, Consumer Product Safety Commission and U.S. Department of Commerce responded that they were looking into acquiring or had recently acquired programs to use AI but gave less details about the technology.
AI capabilities of e-discovery tools
The use of e-discovery platforms that support machine learning features appears more widespread than any other AI use cases in MuckRock’s crowdsourced data.
Lauren Harper, the Daniel Ellsberg Chair on Government Secrecy at Freedom of the Press Foundation, said that agencies are using the term AI very broadly when they refer to these tools as artificial intelligence. The tools may filter out responsive and non-responsive documents within a search, but that can be overridden by a human who’s involved in the review process and not a machine making decisions.
Agencies employ different tools, Harper said, and even if they use the same third-party software, some agencies may pay for features others don’t.
Tools like FOIAXpress, or others mentioned in the reports like NUIX, Relativity and even some functions of Adobe Acrobat, handle tedious parts of fulfilling FOIA requests, like removing duplicate records or news articles. FOIAXpress advertises its ability to “rapidly filter, deduplicate, rank, and sort documents” before “for AI-assisted redaction and processing.”
The simple types of redacting that a tool like FOIAXpress can tackle, like finding all phone numbers across documents, can be very efficient, said Marshall, the Reporters Committee for Freedom of the Press attorney.
But when a task like redaction requires not just identifying but understanding the context of information, Marshall doesn’t think AI should be making decisions. Some redactions require even more information and context, like a foreseeable harm analysis or weighing government interest against public interest. Even when AI can be taught to do that, it may be basing its decisions on agencies that have over-applied redactions or withheld more information than the law permits.
“I think there needs to be a really serious conversation, not just within government, but between government and the requester community, journalists, civil society, about how these tools are going to operate,” Marshall said. “And I don’t think that that has happened yet.”