Further Thoughts on FOIA’s Future: How Might Agencies Use Gen AI to Improve FOIA Processing?
Jason R. Baron
The recent emergence of generative artificial intelligence (“gen AI”) in the form of large language models (“LLMs”) such as ChatGPT has attracted great attention for its use in a wide variety of contexts. Government agencies have been understandably cautious in authorizing employee use of gen AI applications (see OPM Guidance), although some agencies are taking initial steps in the form of pilot programs or studies to determine how gen AI might best be used (FedScoop 2024). To date there has been only modest attention paid to incorporating gen AI into agency FOIA workflows for the purpose of improving FOIA processes. These are, however, very early days, and so I wish to stake out a position that any discussion of how FOIA processes can be improved in the future should involve consideration of gen AI along with other more traditional forms of AI.
Let me start by acknowledging that the well-known resource constraints and institutional inertia that FOIA staff confront every day, especially in the face of growing backlogs of requests, all constitute serious obstacles to curing systemic weaknesses in FOIA processing through AI or other means. Moreover, the idea of adopting the use of ChatGPT or other similar software anywhere in the FOIA process may seem “Star Wars”-ish, given the overall state of AI technology used in FOIA offices to process requests. In particular, although the legal community has been well-versed in the use of “technology assisted review” in large document productions for over a decade (The Sedona Conference, 2019), with only some exceptions most agencies seemingly remain stuck in the era of keyword searching (Muckrock 2024). As I have previously argued elsewhere, this in turn substantially contributes to tremendous delays in agency processing times for review.
I should also say at the outset that the earliest use cases of gen AI in connection with public access requests have taken on a decidedly negative bearing, consisting of efforts on the part of a small number of requesters in “flooding the zone” with AI-generated requests—which in turn have overwhelmed the limited resources of access professionals (Wired 2024). This emergent phenomenon and a variety of other real concerns about the use of AI in FOIA processing are addressed in an excellent paper by Nicholas Wittenberg. In a similar vein, I am on record as expressing skepticism about the hype surrounding the use of gen AI methods in e-discovery, in the absence of rigorous evaluative comparisons with other forms of AI.
With all of these important caveats, in the spirit of Sunshine Week I wish to suggest two ways in which I believe progress can be made towards improving agency FOIA processing using gen AI. First, agencies can achieve a greater level of engagement with requesters using more sophisticated forms of intelligent assistance. And second, agencies using gen AI may someday soon be able to provide better narratives to requesters, explaining the details of how requests have been processed, and most importantly, why documents have been withheld in whole or in part under various FOIA exemptions. I am fully cognizant that neither of these proposals are capable of being implemented today using commercial software out of the box. My hope is that in forums such as these during Sunshine Week a discussion can be had among interested parties in government, academia, and private industry about using gen AI in positive ways to assist in solving FOIA processing issues.
Using Chatbots & Gen AI Prompts to Enhance Agency Engagement with FOIA Requesters
To state the obvious, one of the greatest flaws in current FOIA processes at least at the federal level are the long delays experienced by many FOIA requesters in receiving substantive responses to their requests, beyond initial agency acknowledgment letters. To partially address this problem, the 2022-2024 FOIA Advisory Committee recommended that agencies proactively invite requestors to discuss their requests with agency staff, for purposes of clarification and potential narrowing of scope to enhance processing times and resources spent. This recommendation follows from past DOJ guidance suggesting that as a matter of best practices agencies engage in dialogue with requestors, especially in cases of complex requests.
In the future, the Committee’s recommendation could be implemented including through the use of combined forms of intelligent assistance. In recent years public sector agencies at all levels of government (federal, state and local) have increasingly adopted chatbots and intelligent virtual assistants in a variety of settings as a means of fostering the public’s knowledge of government services in general, as well as to help answer specific questions posed by users of online resources (Digital.gov 2021). “Chatbots are more friendly and attractive to users than, for example, the static content search in frequently asked questions (FAQs) lists. They offer users comfortable and efficient assistance when communicating with them; they provide them with more engaging answers, directly responding to their problems” (Adamopoulou & Moussiades, 2020). At a minimum, government agencies should consider making greater use of traditional rule-based chatbots to handle inquiries regarding FOIA requests already on file.
As distinct from chatbots, gen AI software holds out the potential to greatly enhance the quality of agency dialogue with requesters in a variety of ways. “Diverging from traditional chatbots limited to fixed responses, ChatGPT exhibits a dynamic repertoire, capable of swiftly addressing inquiries, summarizing documents, composing essays, and generating comprehensive content[.]” (Kim 2024). On the other hand, while gen AI software allows for follow-up prompts by users, in its most widely used current commercial forms the software does not accommodate interactive queries in the same fashion as chatbots. A growing field of research exists, however, on the subject of incorporating LLM capabilities into chatbots, producing the best of both worlds in terms of “hybrid” forms of intelligent assistance (Dam 2024).
With respect to FOIA, one can conceive of a day where, subject to proper controls, hybrid forms of chatbots combined with gen AI software could be used by agencies for the purpose of engaging in a dialogue with FOIA requesters, including for the making of requests, as well as clarifying and narrowing FOIA requests. Through the use of gen AI, a requester might have the benefit of understanding in much greater depth what types of agency records exist and where they may be found, based on natural-language descriptions provided. The software could describe in detail what kinds of records have been released previously, with references given to agency online reading room resources. Incorporating chatbot capabilities, users could engage in iterative, back and forth dialogue, for the purpose of understanding what timeline the agency would be operating under to process a given request, and how the agency perceives the reasonableness of requests in terms of their specificity and scope. Ideally, the use of hybrid chatbot and gen AI assistance would mitigate the concern that proactively inviting dialogue with requesters will strain already taxed human resources.
Using Gen AI Narratives to Improve the Quality of FOIA Determination Letters
The 2022-2024 FOIA Advisory Committee also crafted a model determination letter for agencies to use, with the aim of providing more concrete information to requesters with respect to agency processing of their request. In addition to providing greater specifics on the process by which agency staff conduct searches for responsive documents, the model letter emphasizes the need to inform requesters of the agency’s reasons for withholding documents in full, as well as in part. This is all important information to the requester, and can be accomplished with minimal additional time spent by agency staff.
Recent research has demonstrated the capabilities of gen AI software to enhance the quality of determination letters still further. As one step towards identifying sensitivities in agency records, the author along with colleagues at the University of Maryland and at MITRE have engaged in several rounds of research aimed at evaluating how well classical AI methods (in the form of machine learning) perform in identifying “deliberative process privilege” exempt passages in otherwise responsive documents to FOIA requests. As an approximation to the kind of deliberative material found in agency records, we used documents created by staff in the Clinton White House that were previously restricted under the Presidential Records Act (“PRA”) for 12 years as containing confidential communications with presidential advisors. These records are now open to the public, since under the PRA, FOIA’s Exemption 5 is not a basis for withholding presidential records. The results of the research have shown that classical AI methods are capable of identifying deliberative process material about 70% of the time. These findings are generally in line with the success of technology assisted review methods in connection with parties’ responses to requests to produce relevant documents in litigation (Grossman & Cormack 2012).
In an extension of our research, a further study was conducted to evaluate how well ChatGPT software performed in conducting searches. For our purposes here, the relevant part of this latter study involved creating a variety of ChatGPT prompts to have the software explain why deliberative process passages had been withheld, including a prompt specifically asking for case law citations in support of decisions to do so.
The research showed that even ChatGPT in its prior 3.5 version produces narratives that are at least on par with, and in some cases exceed, the kinds of explanations given in typical FOIA agency final determination letters. These gen AI narratives set out a definition of what constitutes the deliberative process privilege, citing to relevant case law, and in many cases went on to describe at a general level of detail the nature of that process. Such details included the fact that conversations proceeded from lower-level staff to senior management, as well as the nature of the recommendations, options, or proposals contained in the document. Concededly, the narratives have a long way to go before reaching the level of sophistication in letters typically crafted by attorneys in response to FOIA administrative appeals (or by lawyers defending FOIA cases in court). Nevertheless, the research supports the idea that agencies would benefit from considering using gen AI narratives to produce more detailed and carefully nuanced responses to requesters. These AI narratives will only become more sophisticated over time using the latest versions of ChatGPT, Gemini, Claude, LLaMA, and whatever the future holds in this explosively growing area.
As a side note, we did not encounter hallucinations in ChtGPT’s citations to case law of the type that have vexed certain members of the legal community. Nevertheless, aside from human review of the exemption decisions being made by AI software, any deployment of gen AI software in connection with generating narratives in response to FOIA requests would need to ensure that there has been some form of human review of the content of narratives, including presumably legal staff reviewing the accuracy of cited case law.
Critics have expressed legitimate concerns about the black box nature of AI generally, and have argued for greater transparency in how agencies use AI, and now gen AI. As stated in a subcommittee report to the 2022-2024 FOIA Advisory Committee, consideration should be given to “the issuance of guidance to agencies on the subject of disclosure to requesters as to how and when forms of AI have been or are being used, including for (i) conducting searches for responsive records; (ii) redacting exempt material for the purpose of withholding records in whole or in part, (iii) generating response letters to individual FOIA requests; and (iv) any other purposes.” That report went on to recommend that such guidance could consist of annual agency reporting to DOJ, generic notices on agency FOIA websites, and/or “by way of a statement in agency responses to individual FOIA requests.”
It needs to be emphasized that neither AI nor gen AI is a panacea for the future processing of FOIA requests. I have been an evangelist for the past 20 years in advocating the use of AI methods in connection with searches for responsive records in the e-discovery and FOIA contexts. So too, I believe in the prospect that gen AI tools can provide useful “intelligent assistance” to FOIA processing in the ways described here, provided that human review will also be part of any process.
One last point: in a time where the size of government is so much in the news, the future availability of AI should not be used as a rationale for eliminating FOIA staff. At the same time, agencies should not turn a blind eye to the efficiencies and improvements that remain possible through the use of various forms of AI technologies, including the ways in which gen AI can potentially help to bring a measure of greater sunshine to the processing of FOIA requests.
Jason R. Baron is a Professor of the Practice in the College of Information at the University of Maryland. He previously served as a trial lawyer and senior counsel at the Department of Justice, and as the first director of litigation at the National Archives and Records Administration. He has served as co-chair of the D.C. Bar’s E-discovery and Information Governance Committee, and is currently serving his third term as a member of the FOIA Advisory Committee to the Archivist. Among his awards while in public service, Jason is a recipient of the Justice Tom C. Clark Outstanding Government Lawyer Award from the Federal Bar Association. He is a frequent media contributor on recordkeeping controversies, with appearances on CNN, MSNBC, NBC News, Good Morning America, and NPR’s 1A and All Things Considered, and citations in the New York Times, Washington Post, Wall Street Journal, TIME Magazine, Washington Monthly, and numerous other media outlets. He received his B.A. from Wesleyan University and his J.D. from Boston University School of Law.