More AI Records, Harder Comparisons: How Federal Inventories Rewrite Accountability

A growing inventory, an unstable history

The US Federal AI Use Case Inventory expanded from 710 entries in 2023 to 2,133 in 2024 and 3,611 in 2025. Yet more entries did not necessarily make individual systems easier to scrutinize over time.

In their November 2026 preprint, the researchers describe this problem as “palimpsestic accountability.” A palimpsest is a manuscript scraped and written over, with traces of earlier writing remaining. Their argument is that successive inventories similarly rewrite public accounts, leaving readers to reconstruct what changed—and why.

The distinction matters: a missing answer might reflect a removed question, a changed reporting obligation, an agency omission, or a change to the system itself. The latest release alone may not distinguish these possibilities.

Reconnecting records across years

The team studied official Office of Management and Budget files available as of April 2026, alongside reporting instructions, data dictionaries, and relevant policy documents.

Tracking systems required more than comparing identification numbers. The 2024 public file lacked an identifier column; identifiers in the other releases used different formats, and 704 entries in 2025 had none.

The researchers combined exact title matching with similarity scores for titles, descriptions, and bureaus. An assignment algorithm proposed matches among remaining records, and the first author reviewed qualifying candidates. This produced 1,332 matched pairs between 2024 and 2025 and a panel following 150 uses across all three releases. Unmatched entries were not automatically treated as new or discontinued systems.

For answer comparisons, the researchers included only entries required to answer the relevant question in both years. They also distinguished substantive answers, explicit negatives, blanks, unknowns, withheld information, and “not applicable” responses. Requirements could not be reconstructed for 2023 because that release lacked a data dictionary.

Close reading focused on uses subject to greater governance scrutiny. The ADMAPS framework guided attention to organizational processes, algorithmic decision-making, and human discretion, while additional themes emerged from the records.

Three ways disclosure changed

First, the questions changed. The inventories contained 10 fields in 2023, 62 in 2024, and 36 in 2025. The final release removed questions about computing resources, documentation availability, and whether an AI use could take consequential action without direct human involvement. It also introduced questions about operator training and failsafes for deployed high-impact uses.

Some changes improved completeness while reducing comparability. Reporting an AI technique became much more common when the questionnaire moved from free text in 2023 to fixed choices in 2025. However, broad categories such as “classical machine learning” offered less technical specificity than some earlier descriptions.

Commercial AI reporting also moved to a separate agency-level file. That arrangement showed whether agencies used tools for common tasks, but did not provide individual-use accounts of their data or oversight.

Second, classifications controlled disclosure. In 2024, agencies identified uses as rights-impacting, safety-impacting, both, or neither. In 2025, the categories included high-impact, presumed high-impact but not high-impact, and not high-impact. Detailed governance questions applied to deployed high-impact uses.

Of 272 matched uses classified as rights- or safety-impacting in 2024, 44 fell outside the high-impact category in 2025. Sixteen entered the presumed-but-not-high-impact category, which required a justification; 28 went directly to not high-impact, which did not. Eighteen of those 28 provided no explanation.

Required explanations could also be generic. All 32 Department of Justice entries in the presumed-but-not-high-impact category repeated the same sentence invoking the policy definition. These findings establish gaps in the public explanations, not whether particular classifications were substantively wrong.

Third, answers disappeared even when reporting remained required. Among 548 matched uses required to describe their data in both years, 140 gained an answer and 46 lost one. The Justice Department accounted for 105 gains, while Veterans Affairs accounted for 34 losses.

Human oversight remained thinly described. Only 18.1% of deployed high-impact entries in 2025 reported established operator training. Entries often identified a human reviewer without explaining how that person evaluated or overrode an AI output.

Documentation is not operational proof

The study analyzes agencies’ public accounts, not observed practices. A blank does not prove a safeguard was absent, and a completed answer does not establish its effectiveness. Matching could miss substantially renamed or reorganized uses, while the qualitative focus may not represent lower-impact entries. The paper also notes later agency updates that the April snapshot did not capture.

The authors recommend persistent identifiers, preserved revisions, explanations for reclassification, and explicit distinctions between different kinds of missing information. They also call for oversight descriptions specifying reviewers’ evidence, authority, training, and time. These are proposed improvements, not interventions tested by the study.