An independent record-by-record review of all 3,611 disclosures under OMB M-25-21 — May 2026
The Trump administration published the 2025 Federal Agency AI Use Case Inventory in April 2026. We reviewed every record across all 56 agencies. Findings below link directly to individual records; 704 records (19%) — covering HHS, State, TVA, ED, GSA, and SSA — were filed without agency-assigned identifiers, though all are addressable via synthetic row-position IDs assigned by this dashboard.
3,611
Records reviewed
445
High-Impact designations
71.5%
High-Impact records missing required safeguard fields
704
Records with no unique ID
Dataset Overview
Use Cases by Agency — 2025 Inventory (top 25)
High-Impact Classification Rate by Agency — 2025 (agencies with ≥ 10 records, sorted by rate)
AI Category Breakdown: 2024 vs. 2025 (raw record counts)
Development Stage — 2025
Impact Classification — 2025
The High-Impact Framework Is Not Working
M-25-21 replaced broad disclosure requirements with a narrower focus on "High-Impact" systems — defined as AI with significant consequences for individuals' rights, safety, or access to services. The data shows this designation is being applied inconsistently, and in several cases, in apparent contradiction with an agency's own description of the system.
DOT: Zero High-Impact designations across 70 aviation and transportation safety records.
This includes MARS — an AI system being developed to support reduced aircraft separation standards in terminal airspace — and AI used for fatigue risk assessment of air traffic controllers. DOT's own record descriptions invoke safety-of-life language; the impact classification does not reflect this.
NASA: One High-Impact system across 425 records — a physical security camera.NASA-753 (Intelligent Camera Analytics, a facility security system deployed in 2016) is the agency's only unambiguous High-Impact designation. Systems classified Not High-Impact include: Autoresolver (autonomous air traffic separation), AEGIS (autonomous target-selection AI deployed on Mars rovers since 2010), the deep-space astronaut medical assistant (formerly "Doc in a Box"), and INSIGHT (autonomous operations of high-pressure gas systems at a rocket test facility).
SEC: Zero High-Impact designations across 60 records, including deployed enforcement-targeting AI.SEC-34 describes an AI identifying accounts with trading activity preceding material price moves "warranting further investigation." SEC-26 flags accounts for examiner review. Both deployed; neither High-Impact.
Even Correctly Classified Systems Lack Required Documentation
Of the 445 High-Impact records, 71.5% are missing all required safeguard fields: testing conducted, impact assessment, independent review, ongoing monitoring, failsafe presence, and appeal process. The one field with universal completion is public consultation, the least operationally specific. Even where agencies correctly identify a consequential system, the documentation required to evaluate its governance is mostly absent.
Notable Individual Disclosures
HHS deployed AI to scan grant materials for compliance with Executive Orders 14151 and 14168.
Two HHS/ACF systems — one reviewing position descriptions and one reviewing grant applications — explicitly cite EO 14151 ("Ending Radical and Wasteful Government DEI Programs and Preferencing") and EO 14168 ("Defending Women From Gender Ideology Extremism and Restoring Biological Truth to the Federal Government"). Both are deployed on Palantir's ACF Upstream platform (grant materials also uses Credal); both are classified Not High-Impact. A companion HRSA system, NOFO Compliance Assistant, is in pre-deployment and will evaluate Notice of Funding Opportunities against "dynamically-changing Executive Orders (EO) and OMB memos to ensure NOFO compliance with White House priorities."
DOE's ACORN system autonomously controls nuclear reactor processes and is classified Not High-Impact.ACORN (DOE-527) "automatically identifies process models from simulation and operational data, solves for optimal control actions that can achieve user-defined objectives, executes actions and observes system responses" in a reactor context at Idaho National Laboratory. A separate record, the Cognitive Prescreen Tool (DOE-59), has been deployed since 2019 assisting classifiers in determining the classification level of documents — potentially including nuclear weapons design information — and is also Not High-Impact.
The VA is applying acoustic-linguistic AI to Veterans Crisis Line calls.
Record VA-25-5182 describes a pre-deployment system processing crisis hotline audio using "linguistic, acoustic, and contextual features" to score suicide risk, also incorporating external environmental data such as noise pollution. It is classified High-Impact. The disclosure does not describe whether callers are notified or what safeguards govern the scoring outputs.
PATTERN, the First Step Act recidivism tool, appears twice under the same record ID with conflicting impact classifications.DOJ-0160 appears once as Not High-Impact and once as High-Impact in the published data. PATTERN scores directly affect early release eligibility and programming placement for approximately 150,000 people in federal custody. The BRAVO Classification system (DOJ-0146) uses statistical prediction to assign security levels to newly admitted inmates before any conduct has occurred inside a facility; it is classified High-Impact with no disclosure of model inputs, training data, or demographic disparate-impact analysis.
State Department retired three geopolitical AI systems — "Violence Against Civilians Model," "Mass Mobilization Model," and "Senturion Alpha" — with blank descriptions.
All three are listed as Not High-Impact, Retired, with no description, no rationale, and no operational history in the 2025 submission. All three were operated by the Bureau of Conflict and Stabilization Operations (CSO). The 2024 inventory — where all three appeared as active systems — provides context: Violence Against Civilians Model "forecast mass civilian killings for the upcoming quarter and year for each country globally"; Mass Mobilization Model predicted "mass mobilizations (protests and riots) for each country globally"; Senturion Alpha "identifies where key decision makers fall on an issue spectrum and who influences whom," simulating political dynamics to estimate influence. The 2025 retirement entries (US25-3935, US25-3936, US25-3937) contain only the system names.
Multiple Major Agencies Filed Near-Blank Submissions
Commerce filed 223 records — all "Not Disclosed" for impact, all blank status, 89% with no description — including NOAA's hurricane track models and flood prediction AI. NIST, the agency that writes the federal AI Risk Management Framework, disclosed three use cases consisting only of cloud vendor names. TVA submitted 59 records with blank IDs, blank status, and no impact classifications for a portfolio that includes nuclear plant operations AI and dam safety monitoring.
Note on Citations
Record ID links open a detail card directly on this page. For 704 records (HHS, State, TVA, ED, GSA, SSA) that were filed without unique identifiers, this dashboard assigns synthetic IDs (US25-n, based on row position in the source CSV). These IDs are not official government identifiers. All records are in the OMB's GitHub repository; the raw CSV is also in this project's source repository.