Skip to content
Article reader Listen + reading controls
LISTEN + READ YOUR WAY

Article reader

Preparing the reader…

0:00 0:00
Reading settings
Text size
100%

AI will not fix FOIA until agencies fix the records

By the time a Freedom of Information Act request reaches an analyst, the hardest problem may already be years old.

The relevant records may be scattered across email, shared drives, case systems, collaboration tools, contractor environments, and personal filing habits. The request may cross several components with different search practices. Reviewers may need to reconstruct context that was obvious when the work occurred but disappeared when people moved on.

That is the operating reality behind the 2024 NextGen FOIA Tech Showcase. The Chief Freedom of Information Act (FOIA) Officers Council invited vendors to demonstrate artificial intelligence (AI), electronic discovery, search, case-processing, and redaction tools that might improve federal disclosure work.

The technology can help. It cannot repair missing records, unclear ownership, inconsistent retention, or a workflow no one has measured from end to end. If agencies treat FOIA as a document-processing problem that begins when a request arrives, they will automate the visible end of a much larger knowledge-management failure.

Revised and substantially expanded July 17, 2026, using later government-wide FOIA data and oversight findings.

Demand is growing faster than the old model

Federal agencies received a record 1,501,432 FOIA requests in fiscal year 2024, a 25 percent increase over the previous year. Agencies also processed a record number—nearly 1.5 million—which shows that the workforce was not standing still. Yet government-wide backlogs and the age of old requests remain persistent problems.

The U.S. Government Accountability Office (GAO) examined the backlog problem in a report released days after the original article. It found that staffing constraints, growing volume, request complexity, records-management challenges, and interagency consultations all contribute to delay. GAO's recommendations on backlog management emphasized goals, milestones, metrics, and reliable data—not a single technological remedy.

This matters because “FOIA automation” can imply that the expensive part of the process is a repetitive task waiting for a model. Some work does fit that description. Much of it involves finding the correct universe of records, interpreting scope, applying law to context, coordinating among offices, protecting legitimate interests, and creating a defensible account of what was searched and withheld.

Tools should be inserted into that system with precision.

Where technology can help

There are several promising uses for AI and adjacent technologies in FOIA operations.

Request intake and clarification. A well-designed service can help requesters describe subject, timeframe, offices, custodians, and desired formats. It can suggest existing public records before a duplicate request enters the queue. It should not silently narrow a request or create the impression that a requester must use particular legal language.

Record discovery. Search, entity extraction, semantic retrieval, deduplication, email threading, and near-duplicate detection can help analysts find relevant material across large collections. The value depends on the completeness of the indexed sources and the ability to explain what was searched.

Triage and routing. Systems can identify likely custodians, components, consultations, complexity, and duplicate requests. Routing should remain observable so a wrong classification does not bury a time-sensitive request in the wrong queue.

Review support. Tools can identify personally identifiable information, privileged patterns, classified markings, and passages similar to prior decisions. These are candidate flags, not final legal conclusions. The context that determines an exemption may not be visible in the text alone.

Redaction quality. Software can assist with consistent redaction across duplicate or related records and check that underlying content has actually been removed from released files. Human reviewers still need to confirm both the legal basis and the technical result.

Proactive disclosure. Repeated requests can reveal information that should be published once in an accessible form. Analytics can help agencies identify those patterns and measure whether publication reduces later demand.

None of these uses requires an autonomous system to make disclosure decisions. In many cases, the best design is narrow assistance attached to an accountable human workflow.

The search problem begins with records management

A retrieval system can search only what it can access and interpret.

Agencies need to know which repositories hold responsive federal records, who owns them, what retention and legal-hold rules apply, how contractor records are obtained, and whether file formats and metadata remain usable. Modern collaboration systems complicate the unit of a record: a decision may be distributed across a chat thread, linked document, meeting transcript, task ticket, and later edit history.

If a tool searches email well but cannot reach the case system where the decision was implemented, high relevance scores create false confidence. If two components use different names for the same program, a literal search may miss one. If records were never captured, no model can infer a defensible release set from what remains.

FOIA performance is therefore downstream of knowledge infrastructure. Records schedules, authoritative repositories, metadata, access controls, system inventories, and contractor clauses may feel remote from disclosure work. They determine its cost and credibility.

Preserve the evidence behind the response

AI-assisted review introduces a new accountability requirement: the agency must be able to reconstruct how the tool influenced the result.

For each request, the record should preserve:

  • the scope and any clarification agreed with the requester;
  • repositories, date ranges, custodians, and search methods used;
  • tool and model versions and relevant configuration;
  • items included, excluded, deduplicated, or routed for additional review;
  • human decisions that accepted or overrode tool suggestions;
  • exemptions applied and the authority for them;
  • consultations, referrals, and queue time at each stage; and
  • the final release package and any later correction or appeal outcome.

This does not mean logging every internal token processed by a model. It means preserving enough decision provenance to evaluate the adequacy of the search, explain the response, investigate failure, and improve the process.

A black-box claim that “the AI found no additional responsive records” is not a defensible search methodology.

The temptation to automate final disclosure decisions will grow as language models become better at reading large record sets and explaining apparent patterns. The legal and institutional risks remain substantial.

An AI system can produce a confident but invented rationale. It may apply a pattern learned from prior redactions after the underlying policy changed. It may miss context outside the document or reproduce inconsistent historical practice. It can also make systematic errors at a scale that manual sampling does not immediately reveal.

Human review is not valuable merely because a person clicked approve. Reviewers need access to the source record, the suggested action, the legal basis, relevant context, and a practical way to disagree. Agencies should test whether automation bias causes people to accept machine recommendations more readily than equivalent human suggestions.

Authority should remain with identifiable officials. The system may prioritize, retrieve, compare, flag, summarize, and check. People remain responsible for interpreting the request, applying exemptions, conducting segregability analysis, and issuing the agency's determination.

Measure the whole flow of work

The most misleading FOIA technology metric is the number of pages “processed by AI.” It counts machine activity, not public access.

Better measures include:

  • time from receipt to perfected request;
  • queue time and working time at each stage;
  • search rework caused by missing sources or unclear scope;
  • age and complexity distribution of the backlog;
  • consistency and error rates in redaction suggestions;
  • appeal outcomes and search-related remands;
  • time saved per request after human verification;
  • accessibility and usability of released records; and
  • demand avoided through proactive disclosure.

The Justice Department's Office of Information Policy (OIP) now publishes quarterly and annual performance data, and its 2025 backlog-reduction guidance stresses adaptable plans, senior-leader involvement, milestones, and metrics. Technology should help an agency see and improve that flow. A faster redaction step will not reduce response time if records collection or consultation remains the binding constraint.

Treat FOIA as a public service and a learning system

The NextGen showcase was useful because agencies need to understand what modern tools can and cannot do. Vendor demonstrations should be the beginning of evaluation, not evidence of readiness.

Before acquisition, agencies should test products with representative record types, exemptions, languages, accessibility needs, and failure cases. They should negotiate access to logs, exports, performance evidence, configuration, and exit support. They should involve FOIA professionals, records officers, technologists, security and privacy staff, counsel, and actual requesters in the design.

Most important, they should use what FOIA reveals about the organization. Repeatedly difficult searches identify weak records practices. Recurring consultations reveal unclear authority. Common requests point toward proactive publication. Appeals expose where reasoning or communication needs improvement.

AI can help agencies find and review records. The larger opportunity is to make government information easier to manage, explain, and release before another backlog forms around it.

Sources

READER-NEUTRAL SUBSCRIPTION

Follow Field Notes via RSS.

Copy this address into the RSS reader you already use. New notes will appear there automatically—no account, email address, or tracking required.

Open the feed