Palimpsestic Accountability in the US Federal AI Inventories

Dipto Das, Syed Ishtiaque Ahmed, Aditya Vashistha, Shion GuhaarXiv 2026 · arXiv

TL;DR

By tracking the US Federal AI Use Case Inventory across three years (2023–2025), researchers discovered that government accountability is often "rewritten" rather than simply updated. They propose the concept of palimpsestic accountability, where successive releases of public records overwrite previous ones—often removing critical details about human oversight and risk—leaving the public to "re-stitch" the history of how AI is actually being governed.

What Problem Does This Paper Address?

Governments worldwide are increasingly using AI inventories—public lists of AI systems—to demonstrate transparency and accountability. However, most research treats these inventories as static snapshots. In reality, these are living documents.

The authors identify a major gap: we don't know how agencies change their stories about the same AI system over time. Because these inventories often lack persistent identifiers (like a unique ID that stays with a system year-over-year), it is nearly impossible for the public to track if a system’s risk level was downgraded, if human oversight was removed, or if its technical description was sanitized. The paper tackles the challenge of longitudinal transparency in public-sector AI.

Methods

The researchers conducted a longitudinal mixed-methods study of the US Federal AI Use Case Inventory releases from 2023, 2024, and 2025.

  1. Record Linkage: Since the 2024 inventory lacked IDs and names changed frequently, the authors used a mixed-method approach to link entries. This included exact matching on titles/identifiers and a "fuzzy" matching process using the Hungarian algorithm to score similarity across titles, descriptions, and bureaus, followed by manual expert review.
  2. Quantitative Analysis: They tracked how many fields were required, how many were left blank, and how classifications (like "High Impact") shifted.
  3. Qualitative Analysis: Using the ADMAPS (Algorithmic Decision-Making Adapted for the Public Sector) framework, they performed a deep dive into "High Impact" cases. They coded how agencies described bureaucratic processes, algorithmic roles, and human discretion.

Datasets

The primary data source was the official US Federal AI Use Case Inventory (consolidated files for 2023, 2024, and 2025).

  • 2023: 710 entries, 10 fields.
  • 2024: 2,133 entries, 62 fields.
  • 2025: 3,611 entries, 36 fields.

The authors have produced a linked dataset of these entries to support future longitudinal research, which they intend to publish upon the paper's formal acceptance.

Key Insights and Ideas

  • The Palimpsest Metaphor: The authors liken the inventory to a palimpsest—a manuscript where the original text is scraped away to make room for new writing, but traces of the old remain. In AI inventories, new releases don't just add information; they often "write over" previous governance accounts.
Figure 1: A palimpsest is a manuscript scraped and written over, with traces of earlier text still visible. One AI use case reported across three releases (2023–2025) of the US Federal AI Use Case Inventory, where each release writes over the last. Annotations show what changed in the entry from the previous release.
  • Three Mechanisms of Rewriting:
    • Redrawing: Each release changed the questions asked. For example, the 2025 release removed questions about computing resources and whether an AI could act without a human, making these constraints invisible.
  • Gating: Disclosure is tied to classification. If an agency reclassifies a system from "High Impact" to "Not High Impact," it is suddenly exempt from reporting on impact assessments and appeals.
  • Overwriting: Even when questions stayed the same, answers changed. A detailed data description in 2024 might be replaced by a blank cell in 2025, even if the system itself didn't change.
  • Agency-Level "Personalities": Accountability gains and losses were often concentrated in specific agencies rather than being a government-wide trend. For instance, the DOJ moved from leaving many fields blank to filling them, while the VA did the opposite.
Figure 3: AI techniques and commercial AI tasks reported in the inventories.

Key Results

  • Explosive Growth, Shrinking Detail: While the number of reported AI use cases grew from 710 to 3,611, the number of questions asked per case dropped significantly in the latest release (from 62 to 36).
  • The "High Impact" Trap: About 1 in 6 (16.2%) AI uses classified as "rights-impacting" or "safety-impacting" in 2024 were downgraded in 2025. Most of these reclassifications were provided with no explanation or a generic template sentence.
  • Human Oversight is a Black Box: Even for "High Impact" systems, only 18.1% reported established operator training. Descriptions of human oversight often used "routing" language (e.g., "sent to a manual queue") without explaining how humans actually exercise judgment or override the AI.
  • Commercial AI Invisibility: In 2025, commercial AI (like Microsoft Copilot) was moved to a separate, aggregate report. This removed use-case-level transparency for the most widely adopted tools in the federal government.

Limitations and Future Work

  • Public Accounts vs. Reality: The study analyzes what agencies say they do, not their actual internal practices. There may be a gap between the public "palimpsest" and the ground truth.
  • Matching Errors: Despite rigorous fuzzy matching and manual review, the lack of persistent IDs means some systems that were renamed or merged might have been missed.
  • Focus on High Impact: The qualitative analysis focused on the most scrutinized systems; lower-impact systems might be governed (or ignored) in very different ways.

Future Work: The authors suggest that future research should involve interviews with the civil servants who actually fill out these inventories to understand the organizational pressures that lead to "overwriting" and reclassification. They also call for the OMB to implement persistent public identifiers to make "re-stitching" these records easier for the public.

Summary updated

Abstract

An AI inventory is a structured, often public record in which a government or organization lists and describes the AI systems it uses or plans to use. Studies of single AI inventory releases describe what systems do but say little about how they are governed over time. We linked entries across three releases of the US Federal AI Use Case Inventory (2023--2025), compared answers only where both releases required them, and analyzed governance accounts through a dual deductive-inductive approach, guided by the Algorithmic Decision-Making Adapted for the Public Sector (ADMAPS) framework. As the inventory grew from 710 to 3,611 entries from 2023 to 2025, fewer required answers were left blank, but accountability was rewritten through three mechanisms. Each release \emph{redrew} the scope of the questions; agency-assigned classifications \emph{gated} what was reported; and answers were also \emph{overwritten}, with gains and losses concentrated in single agencies and little account of human oversight and judgment. These findings show that each release redraws the boundary of accountability and leaves the public to \emph{re-stitch} the records, a condition we term \textbf{palimpsestic accountability}. We recommend versioned records, reasons for every reclassification, public files that distinguish kinds of missing answers, and oversight documented as conditions to improve public sector AI accountability.
BibTeX
@misc{das2026palimpsestic,
  author = {Dipto Das and Syed Ishtiaque Ahmed and Aditya Vashistha and Shion Guha},
  title = {{Palimpsestic Accountability in the US Federal AI Inventories}},
  year = {2026},
  eprint = {2611.04153},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.04153},
}
Figure