A few federal agencies have begun to use sophisticated artificial intelligence systems to assist with massive caseloads of Freedom of Information Act requests, but a few transparency advocates caution that the government needs extra protections before adopting the technology more generally.
In current interviews, federal officials verified that at least three agencies — the State Department, the Centers for Disease Control and Prevention, and the Justice Department, — have attempted out or are currently testing machine-learning algorithms and models to help in search for information in databases retaining billions of government records.
Officials from other agencies have also separately examined an AI prototype known as “FOIA Assistant” that is being developed by a federally financed research organization as a viable model for handling record-high volumes of new requests and mounting backlogs of current ones.
“There is no methods for FOIA to function in the future unless you automate the finding of the millions, hundreds of millions, as well as billions of records which these government agencies maintain,” said Jason R. Baron, an information studies professor that the University of Maryland and an acknowledged expert on the utilization of artificial intelligence for government access. “The problem seems unsolvable without AI.”
Despite the American legal system, which has utilized court-approved “eDiscovery” technologies over years to help discover and retrieve sensitive information through documents exchanged during litigation, he added, the application of artificial intelligence to FOIA purposes remains in its early stages.
Nonetheless, some open governance and civil rights organizations are already concerned that the government’s approach toward employing AI to assist with FOIA issues may create new ones.
According to Adam Marshall, a senior staff attorney with the nonprofit government watchdog Reporters Committee for Freedom of the Press, AI and other technology will assist in making more information accessible to the public faster. But first, he says, it’s important to understand how technology “is being trained and utilized by humans.”
According to Marshall, government entities have not generally disclosed to the public what types of AI tools are being deployed and in what context. He also expressed concern that overloaded FOIA officers exposed to AI may become overly reliant on or comfortable with robots in making conclusions that generally require careful legal research.
“There needs to be clear standards for the utilization of this technology, as well as assurances that they’re being followed,” Marshall said. “There must also be processes in place for challenging choices made by machine algorithms, including when they may be withholding information unnecessarily or illegally.”
The Freedom of Information Act, signed into law in 1967, is intended to guarantee government accountability and accessibility to information by compelling agencies to give records to citizens who make requests.
However, experts generally agree that the FOIA process has to be modernized and improved, as requests can frequently take months, if not years, to process. A growing number of requesters are turning to the courts for assistance in obtaining records in a timely way.
According to the Justice Department’s Office of Information Policy, the 120 federal agencies that are covered by the federal disclosure legislation received over 928,000 FOIA requests last year, an all-time high over 90,000 more than in 2021. Meanwhile, the total number of backlogged requests in 2022 reached a new high of about 207,000, an increase of nearly 50,000 from the previous year.
Backlogs are growing at a time when agencies are expecting a surge in electronic record storage and disclosure, according to numerous officials.
According to Michael Sarich, that oversees FOIA challenges for the Department of Veterans Affairs and co-chairs the Chief of the FOIA Officers Council’s technology subcommittee assigned with exploring the implementation of AI, a group of FOIA officers has been spreading information on “the awareness and accessibility of these types of tools” as part of an effort to improve the FOIA process.
“The volume of information, records, and data that’s now accessible was unimaginable 56 years ago while the FOIA was created,” said Eric F. Stein, an additional subcommittee co-chair and deputy assistant secretary of the State Department, who oversaw FOIA initiatives, declassification strategies, and other record management planning for the agency.
According to Stein, the State Department is now evaluating two AI models to assist with FOIA requests. According to him, one model uses machine-learning algorithms to discover records in the agency’s centralized databases & archives, which contain more than 3 billion entries.
The other pilot sends prompts to those who submit FOIA requests through the agency’s web portal based on the words they enter, suggesting that they might be able to find data that is already publicly available or helping them narrow the scope for their requests to help facilitate faster responses, he said.
According to a Justice Department official, the use of AI in FOIA processing is now “purely exploratory” and not a policy at the department.
The CDC also “explored using an AI software program to analyze documents submitted into the FOIA system,” according to an NBC News spokesman. “Ultimately, this software did not meet the office’s needs in an efficient and timely manner, and it was determined that the tool was not an appropriate fit (for the) intended use.”
Some federal agencies have begun testing a new prototype called “FOIA Assistant,” which assists in locating records inside massive government databases and offers redactions of information within at least 3 of the law’s 9 exemption categories.
“There’s really no commercially available tool like this one that helps FOIA analysts that we’re aware of,” explained Bradford Brown, the project’s outcome lead at Mitre Corp., a nonprofit manager for federally supported government research and development initiatives that produced the prototype.
Brown stated that Mitre developers collaborated with FOIA analysts to curate datasets to assist train and test the model by detecting and annotating portions of information barred from disclosure because they contain “deliberative language” meant to help authorities make judgments.
Baron, a former litigation director for the National Archives and Records Administration, said he assisted in testing the prototype through annotating numerous Clinton administration policy documents, and discovered that an early version of the system was roughly 70% correct.
“It’s not perfect,” stated Baron. “However, using this type of AI could be extremely beneficial in the future when agencies routinely find tens or hundreds of thousands of prospectively responsive records which they would otherwise have to examine manually, a process that will almost certainly take many years.”
Brett Max Kaufman, an ACLU senior staff attorney that specializes in surveillance & national security issues, acknowledged that artificial intelligence models may eventually speed up the disclosure of government information, but he warned that this gain may come at a cost.
“Agencies regularly over-redact & over-withhold information under FOIA,” he stated. “There’s a culture of maintaining things as secret as possible lasting as long as possible, partly because all of the incentives are geared toward that.” And just teaching a computer to perform the same thing you’ve always done is likely to make things worse.”
Brown, who highlighted that the Mitre prototype is still in the works, declined to reveal which of “multiple agencies” has tested it so far. He went on to say that the gadget is only meant to be a “cognitive assistant to help analysts.”
“In the end, humans will still have to make decisions,” he remarked.