حكومة الولايات المتحدة تُصوِّر 3،611 حالة من استخدام الذكاء الاصطناعي - تصل إلى 70٪ من عصر بايدن

3,611 active or planned AI use cases across the U.S. federal government. That's the number the Office of Management and Budget disclosed on April 14, and it represents a 70% increase over the final Biden-era inventory. The list includes systems that would make Philip K. Dick wince: HHS hired Palantir to scan grant applications for ideological alignment, the Bureau of Prisons is building an AI to assess inmate misconduct potential before any actual misconduct, and the VA wants an AI to listen in on veterans crisis line calls and pull external database info to estimate suicide risk.

Palantir, Prisons, and the Bureau of AI Misconduct

Health and Human Services' Administration for Children and Families contracted Palantir - the company known for work with the military, CIA, and ICE - to flag grants it deems misaligned with the administration's policy dictates. The Federal Bureau of Prisons is developing a system to predict which newly admitted inmates might commit misconduct, routing them to high-security confinement before they've done anything wrong. The Department of Energy is testing AI to autonomously control nuclear reactors, targeting responses to potential safety incidents. The State Department? It ended a program that used AI to forecast mass civilian killings, a tool meant for conflict prevention.

Nuclear Reactors and Suicide Hotlines: AI in Life-and-Death Decisions

The VA's crisis line AI pulls data from external databases to assess mental state and suicide risk in real time. DOE's reactor control AI aims to autonomously respond to nuclear safety incidents. These are not fringe experiments; they are active or planned use cases on a government-wide registry. Predictive methods for prisoner classification go back decades, and model predictive control of nuclear reactors is well studied. But the scale and speed of this deployment eclipse prior efforts. The 2025 inventory had 70 machine translation use cases alone, up from 58 in 2024.

Transparency Theater: Why the Inventory Tells You Almost Nothing

Each use case description is typically a single sentence, rarely more than a paragraph. The OMB's disclosure carries minimal context. No impact assessment, no public comment period - only one of the examples cited (DoJ) even proposes public involvement. The rest are not classified as "high impact" under the administration's own policy, a label applied inconsistently. Canada offers a contrast: a federal directive mandating transparent risk-scoring and impact assessment for automated systems making administrative decisions, with stakeholder consultation from conception. France's 2016 Digital Republic Act requires public records access, human appeal, and mandatory notification for all algorithms used in government administrative decisions.

The U.S. should adopt algorithmic impact assessment procedures and public comment processes. The Canadian and French models show it can be done meaningfully. Without that, the 3,611 use cases will remain a black box - and the public's distrust will be fully justified.

Source: AI Use by the US Government
Domain: schneier.com