What the Federal AI Inventory Leaves Unstated

The federal government's consolidated 2025 inventory contains 3,611 agency-reported AI use cases. Only 1,040 are marked deployed; the file also covers systems in pre-deployment, pilot, retired, and unstated stages. Inside that wider register, 445 systems are designated high-impact. For 318 of them, the agency left the pre-deployment-testing field blank. Only 16 report an established appeal process. The silence is a disclosure finding. It is not proof that a safeguard does not exist.

This is an inventory, not an audit

The distinction begins with the source. The Office of Management and Budget's official repository consolidates what individual agencies reported about their own systems under EO 13960 Sec. 5; Advancing American AI Act; OMB M-25-21. It does not independently test those systems or certify the agencies' answers. A record can show that an agency claimed a safeguard was established, in progress, not applicable, or left unstated. It cannot show that the safeguard worked.

Nor is 3,611 a deployed-system count. The source spans the whole reported lifecycle:

The inventory also contains 24 entries whose agencies withheld details. Their presence is countable; the withheld facts are not reconstructed or inferred.

The 445-record denominator

Safeguard analysis belongs only on the 445 records whose designation is exactly high-impact. The source also uses language such as “presumed high-impact, but not high-impact.” Those records are not included merely because the phrase contains the same words. Exact designation matching is the difference between a defensible denominator and an inflated one.

The inventory asks agencies to describe eight safeguards for those designated systems: testing before deployment, an impact assessment, independent review, ongoing monitoring, operator training, a human failsafe, an appeal process, and public consultation. Agency wording is normalized into four comparable states, while the original wording remains in the machine-readable record.

Most safeguard fields are unstated

Across all eight safeguards, between 317 and 318 of the 445 high-impact records contain no agency answer. That is roughly seven in ten records for every safeguard measured.

Every row sums to 445. “Unstated” means the source field was blank; it is deliberately not labeled “no.”

Testing is unstated for 318 systems

On pre-deployment testing, 45 high-impact records report an established safeguard and 82 report work in progress. The remaining 318 — 71.5% of the denominator — say nothing in that field. The register therefore supports a claim about disclosure completeness: most high-impact records submitted by agencies do not state whether pre-deployment testing was established. It does not support the stronger claim that testing never occurred.

An established appeal appears in 16 records

The appeal row is narrower still. 16 records say an appeal process is established, 78 say one is in progress, and 34 classify it as not applicable or precluded. For the other 317, the field is blank. An affected person reading only this inventory therefore gets an affirmative, established appeal answer for 3.6% of the high-impact systems.

That percentage describes the register, not necessarily the underlying administrative process. An agency may provide review elsewhere, may have classified the question differently, or may simply have omitted the answer. Those possibilities are exactly why silence must remain its own category.

The high-impact list is concentrated

The first three agencies in the high-impact count — Department of Veterans Affairs, Department of Justice, Department of Homeland Security — account for 384 of 445 designations (86.3%). That concentration makes agency reporting practice important: a vocabulary choice or a blank-heavy submission from one large reporter can move the federal total.

Counts are agency self-classifications. They should not be treated as a performance ranking or as proof that agencies with fewer designations use less consequential AI.

How to read the finding honestly

• Supported: most high-impact safeguard fields in the consolidated 2025 inventory are blank.
• Supported: only 16 high-impact records affirmatively report an established appeal process in this file.
• Not supported: every blank means the agency had no safeguard.
• Not supported: all 3,611 entries were deployed government AI systems at the same time.
• Not supported: a reported safeguard was independently tested or found effective.

Source, privacy, and reproducibility

The analysis is generated from OMB's machine-readable consolidated file and was last rebuilt on 2026-08-21. The complete high-impact records and an index of all other reported use cases are available in the Federal AI Use Case Inventory, with keyless JSON at /ai-inventory/index.json.

The source's contact column was dropped in full because it included named federal employees and non-government contact data. The public version contains institution-level records only. Fields describing whether a system processes personal data are facts about the system, not personal data about an individual.

Edition: 2025 consolidated inventory. Publisher: US Office of Management and Budget, Office of the Federal CIO. Source basis: US federal government work, public domain under 17 U.S.C. 105 (by authorship — the publisher ships no explicit LICENSE file). Every safeguard count is computed over the exact high-impact designation, and each normalized state retains the agency's original wording for auditability. See the data standards for the site-wide naming and evidence rules.