FOIA Queues Are Filling With AI Requests. Agencies Must Adapt.

Contents9 sections

Cuillier's July Estimate

In July, roughly 29 percent of the FOIA requests arriving at the U.S. Department of Education appear to have been drafted with help from an AI tool. The estimate comes from David Cuillier, who analyzed the text of the requests themselves: "In the month of July, about 29% of the requests sent to the Department of Education were crafted by AI, in part or fully," he told Straight Arrow News. It is a researcher's estimate derived from language patterns, not a figure the agency publishes or endorses, and it should be read that way.

Cuillier directs the Brechner Freedom of Information Project at the University of Florida, founded the peer-reviewed Journal of Civic Information, and serves on the federal FOIA Advisory Committee under the National Archives. He has spent two decades training journalists and citizens to file records requests, which puts him in an unusually good position to recognize when a request was not written by a person sitting down with a single question.

We work with federal, state, and local organizations on the machinery underneath this: intake queues, case management, records governance, and the AI tooling agencies are now evaluating to keep pace. Our interest is operational. Whatever any office believes about AI-assisted requesting, the requests are arriving, and three decisions, how intake is designed, how search scope is set, and how AI materials are scheduled as records, together determine whether that volume gets absorbed or becomes next year's backlog.

Why Requests Are Longer

Close-up of a hand using a yellow highlighter on a dense block of printed text on paper.

The length comparison translates directly into labor. Cuillier found that AI-drafted requests were typically more than twice as long as human-written ones, running up to about 270 words against roughly 130 for ones written unaided, and that most of them, about 83 percent, enumerate many categories of records rather than a narrow set. The consequence for the office receiving one is mechanical: as Straight Arrow News described it, such a request "could require an agency to search through large amounts of records or data across multiple departments or databases".

A language model asked to draft a public records request does what it was built to do, which is to be comprehensive. It names every plausible record type, adds "any and all," widens the date range, and lists adjacent program offices in case the first one is wrong. The requester usually wants one specific thing; what lands in the queue asks for twelve.

Scope clarification is the first control available to a FOIA office, and it resolves ambiguity without requiring a denial. The statute obliges agencies to produce records when a request reasonably describes what is sought and is made in accordance with published rules. As the National Security Archive frames the test, what matters is whether agency staff can reasonably ascertain which records are being requested and locate them. A 270-word request listing eight record types across four offices frequently fails that test, not because the requester was unreasonable but because the drafting tool buried the one thing they actually wanted.

Open-government advocates have good reason to watch how agencies use this. The National Security Archive has raised concern that clarification letters can function as a soft close, with requests shut when a requester does not respond fast enough, and it has proposed that agencies allow a minimum of 30 working days to supply what is needed. Clarification done well narrows the search and gets the requester what they came for. Used as attrition, it produces appeals and litigation that cost more than the search would have. The difference sits in process design: a standard template, a documented response window, and a tracked disposition for every clarified request.

When an AI Prompt Becomes a Federal Record

An open laptop showing a chat interface, on a wooden desk beside stacks of file folders and rows of metal filing cabinets.

While requesters point AI at agencies, agencies are pointing AI at their own work, and that produces material someone will eventually ask for. The National Archives and Records Administration (NARA) addressed this directly in AC 11.2026, a memorandum issued to federal agency records officers on August 21. The guidance comes in two parts, the first explaining "how to apply the definition of a federal record to inputs, outputs, data, audit trails, software, and other materials involved in the use of AI," and the second confirming "that agencies may only dispose of AI-related federal records in accordance with a NARA-approved records schedule".

The qualifier carries more weight than the headline. Agency use of AI does not automatically manufacture federal records. NARA ties the determination to the circumstances surrounding the creation, maintenance, and use of the materials, and scopes the memorandum to records management under the Federal Records Act rather than AI governance, privacy, security, or ethics. Records management software vendor QFlow Systems, reading the memo for its own customers, offers one plausible reading of where that line falls: an audit trail a staffer switches on for personal convenience, never circulated and never used to make a decision, may be a personal file rather than a record, though NARA's text does not spell out that specific example itself. The same prompt captured in an agency system, retained, and used to support an action is a different thing entirely.

What this asks of agencies now is a capture decision. Which AI tools in the environment have their prompts and outputs retained, under which schedule, and in a system the FOIA team can search when a request asks for them? Offices that have not answered that question tend to find the material sitting in a vendor's logs, unscheduled and invisible, at the moment they are obliged to say what exists.

This runs alongside the governmentwide AI policy framework rather than separately from it. The Office of Management and Budget's M-25-21 and M-25-22 are described by legal analysts as requiring agencies to publish AI strategies, inventory use cases, apply minimum risk management practices to high-impact AI, and publicly report risk determinations and waivers with justification. Records capture and AI inventory work are one exercise viewed from two angles. Agencies that staff them as two projects pay twice.

The Backlog Underneath the Surge

An overflowing in-tray stacked with manila file folders on a cluttered office desk, lit by daylight from a window.

Agencies received 1,707,197 FOIA requests in fiscal 2025, a 13.7 percent increase over the prior year, and processed 1,635,055, a record. The backlog still grew to 339,671 requests, up 27 percent, which is about two and a half months of processing at the rate agencies cleared work that year, with nothing new arriving. DHS, the Department of Homeland Security, reported a FOIA backlog of just over 168,000 requests at the close of fiscal 2025, while holding that backlog to 16 percent of requests received against a 19 percent increase in intake; other tallies of DHS's total pending caseload, which include requests still within their statutory response window alongside the backlog itself, run higher, near 245,000.

Staffing loss accounts for much of that increase. Federal News Network's review of Chief FOIA Officer reports found the Defense Department's backlog rose 42 percent past 30,000 cases amid a 37 percent loss or turnover in FOIA officers across components, while the office at the Department of Housing and Urban Development (HUD) "lost 40% of its staff, including contractor support, while also navigating an 80% increase in requests since fiscal 2024". More than 600 government information specialists have left federal service since the start of the current administration. The Government Accountability Office (GAO) has been tracking the climb for a decade.

AI-assisted volume compounds a strain that was already there, and the two problems, rising intake and shrinking staff, take different remedies. An office whose requests are broad but whose throughput per analyst is sound can buy its way to improvement with intake classification, deduplication, and better search. An office that lost 40 percent of its reviewers, as HUD's did, cannot, because nothing in the toolchain signs an exemption determination. Knowing which condition you are in separates a procurement that pays for itself from one that adds a license fee to the same backlog.

Where DHS Is Already Using AI

DHS lists RelativityOne, an e-discovery platform, in its consolidated AI inventory, with a privacy threshold analysis identifying the software as a tool used to timely respond to FOIA requests and the department planning to use it to streamline identification and redaction of sensitive content in FOIA documents. The use case sits in pre-deployment and is categorized as not-high-impact, and the inventory discloses nothing about the risk management practices applied.

That last gap is where we draw our own line. AI earns its place in a FOIA office at triage, intake classification, duplicate detection, and search across large record sets. Release and redaction decisions stay with trained reviewers, because an exemption determination is a legal judgment about a specific record in a specific context. Abigail Kunkler, a law fellow at the Electronic Privacy Information Center, put it plainly to FedScoop: "AI simply cannot make the kinds of determinations that FOIA requires agencies to make". Peer-reviewed work by Eric Robinson and Erin Coyle in the Journal of Civic Information examines the pilots now underway against the statutory, regulatory, and case-law requirements for searches and responses, and proposes protocols for keeping them compliant.

Testing Before Going Live

Before any tool goes near a live queue, an office needs three baselines: average request size measured in words and record categories, time to first substantive response, and the share of incoming requests that require clarification. Those numbers tell you whether scope or throughput is driving your backlog, and they are the only honest way to know afterward whether a tool helped.

Then test on history. Closed requests from prior years are effectively a labeled dataset: the responsive set is known, the search terms are documented, the outcome is final. Running AI-assisted search and deduplication against those closed matters, with reviewers scoring recall and precision against what the office actually produced, carries no statutory risk and shows what the tool does on your records rather than on a demo corpus. The techniques are well established. E-discovery services firm CDS Legal, which sells these capabilities, describes a practice of using deduplication, email threading, near-duplicate analysis, and conceptual search for years to cut the volume of potentially responsive records and the staff hours needed to review them. Jason Baron, a University of Maryland professor and FOIA Advisory Committee member, has argued that technology-assisted review "can help streamline searches, reduce human error, and respond more efficiently to complex Freedom of Information Act (FOIA) requests".

Measurement before go-live keeps an agency defensible when a requester, an appeals authority, or a court asks how a record was found or why it was withheld. GAO has already flagged that many agencies have reported inaccurate processing times and that backlog reduction plans routinely omit key elements. Layering an untested tool on unreliable measurement produces faster processing nobody can account for.

What Spruce Brings to Intake

Our own FOIA solution, which we build and sell to agency clients, covers request intake, tracking, and reporting, including custom reporting built for compliance and oversight requirements. We pair it with records governance and AI advisory work for the same clients, because each recommendation above depends on infrastructure already being in place.

Scope clarification works only if intake captures the original request text, the clarification sent, the response window, and the disposition, in a form the office can report on later. Capture for AI materials is the same discipline one layer down. NARA's requirement that AI-related federal records be disposed of only under an approved schedule presumes someone has already decided which systems hold those prompts and outputs and where they sit in the file plan. That mapping belongs with the records officer and the IT lead together, before a request arrives asking for the prompts a FOIA team used to run a search. Pre-deployment testing, in turn, works only if the baselines were taken first, which means the case management system has to count the things worth counting from day one. Our published position on the dividing line is the one above: AI for triage and search, people for judgment.

What Offices Can Do Now

State and local custodians face the same arithmetic with fewer staff and tighter clocks. Most states require fulfillment within 20 days at most, while federal responses have averaged around 49 days. Legislatures in several states are weighing their own responses to request volume, including proposals targeting high-volume commercial requesters, and those debates will resolve on their own schedule. No office should plan around the outcome.

Volume will keep rising either way. As Federal News Network has observed, AI and automation let requesters generate hundreds or thousands of requests, expanding participation and increasing volume from earnest and abusive requesters alike. Intake design, records capture under NARA's scheduling requirement, and tooling tested against closed requests are available to an office today, on its own authority. The agencies treating this as a design problem will be the ones still meeting their deadlines two years from now.

Sources

Back to Blog

FOIA Queues Are Filling With AI Requests. Agencies Must Adapt.

Cuillier's July Estimate

Why Requests Are Longer

When an AI Prompt Becomes a Federal Record

The Backlog Underneath the Surge

Where DHS Is Already Using AI

Testing Before Going Live

What Spruce Brings to Intake

What Offices Can Do Now

Sources

More articles

Want our take on your AI roadmap?

Related services from Spruce

AI Advisory

AI Roadmap

Cybersecurity Services