On April 14th, the Trump administration did something it almost never does voluntarily: it told the public something. The Office of Management and Budget disclosed 3,611 active or planned AI use cases running across the federal government, a 70% increase from the final year of the Biden administration. The catch, and there is always a catch, is that almost none of those disclosures contain enough information to understand what the hell is actually happening.
3,611 Use Cases, One Sentence Each
The Guardian reports that the OMB inventory was published so quietly that you would only have stumbled across it if you follow FedScoop religiously or monitor the OMB's federal chief information officer's GitHub account. That is not a joke. That is the official transparency mechanism for one of the largest expansions of automated government decision-making in American history.
The descriptions in the inventory are typically a single sentence long. Rarely more than a paragraph. What matters, as researchers Nathan E. Sanders and Bruce Schneier write in The Guardian, are the details of how these systems are used. Those details are almost entirely absent. This is a list of programs, not an explanation of them. It's the government handing you the index of a book it refuses to let you read.
What They Are Actually Doing With This Technology
Let's go through some of the highlights, because they deserve to be read slowly. The Federal Bureau of Prisons is building an AI system to assess the "potential for misconduct for newly admitted inmates" and route people into high-security confinement before they have done anything wrong in custody. That is not a metaphor for something. That is a literal pre-crime classification system running inside American prisons.
The Department of Veterans Affairs is developing an AI that listens in on calls to the veterans crisis line and then pulls data from external databases to assess the mental state and suicide risk of the person calling. A veteran in crisis, reaching out to a human being for help, gets an algorithm quietly running in the background deciding how serious they are. The Department of Energy is testing AI to autonomously control nuclear reactors and respond to potential nuclear safety incidents. And the State Department quietly ended a program that used AI to forecast mass civilian killings for conflict prevention purposes. That last one they just turned off. Gone.
The Health and Human Services office of administration for children and families hired Palantir, the data analytics firm with deep ties to the CIA, military, and ICE, to scan grant applications and flag those not ideologically aligned with the administration's priorities. According to The Guardian, Sanders and Schneier describe Palantir as the world's "scariest AI company." HHS handed them the power to filter which children's programs get funding based on political alignment.
To Be Fair, Some of This Isn't New or Inherently Evil
Here is where intellectual honesty requires a small pump of the brakes. Sanders and Schneier make clear in The Guardian that not every item on this list is a harbinger of dystopia. The use of predictive tools to assign prisoner security classifications goes back decades, even if those systems have a long documented history of racial bias and inaccuracy. Autonomous systems for model predictive control of nuclear reactors are a well-studied field, and The Guardian notes the recently disclosed AI addition was actually initiated under the Biden administration.
The inventory also includes 70 machine translation use cases, up from 58 under Biden. Customs and Border Protection uses AI translation when human interpreters aren't available. Is that perfect? No. Is an officer with a functional AI translator better than one who cannot communicate with the person in front of them at all? Probably. The existence of genuinely benign or even helpful AI use cases does not excuse the rest of it. It just means the problem is not AI in government per se. The problem is what specific AI tools are being used for, by whom, with what oversight, and with what accountability. On all four of those questions, the current inventory offers almost nothing.
The Public Consultation That Isn't
The OMB inventory theoretically involves some form of public consultation. In practice, according to The Guardian's reporting, there is generally none. Only one of the programs cited in Sanders and Schneier's piece even proposed involving the public in its review process.
The rest are exempt because they are not classified as "high impact" use cases, a label that The Guardian reports is applied inconsistently across agencies. So the Bureau of Prisons pre-crime system does not meet the bar for high impact. Listening to veterans in crisis does not meet the bar. Nuclear reactor automation does not meet the bar. What would? The administration has not said. The standard exists to create the appearance of oversight while systematically ensuring that oversight never actually applies.
What Meaningful Transparency Would Actually Look Like
Sanders and Schneier make a straightforward argument in The Guardian: disclosure of AI use cases could build genuine public confidence, but only if it comes with consistent, meaningful public consultation. Washington D.C. and California are both actively engaging the public to determine where and how AI is appropriate in government processes. The federal government, which operates at orders of magnitude larger scale, is publishing GitHub commits.
This is not a technical problem. The federal government knows how to hold public comment periods. It knows how to write detailed program descriptions. It knows how to define impact classifications consistently. These are choices. The choice being made here is to technically comply with disclosure requirements while structuring those requirements so that nothing meaningful is actually disclosed. Call it transparency theater. The curtain goes up, the stage is empty, and the administration takes a bow.
The Dingo Take
Here is the uncomfortable truth at the center of this story. The Trump administration did not have to publish this inventory at all. They did it because there are rules requiring them to. And then they published it in a format so sparse, in a location so obscure, with so little contextual information, that it functionally defeats the purpose of the rule. This is what regulatory compliance looks like when you are being run by people who view regulations as annoyances to be technically satisfied rather than obligations to be meaningfully honored.
The 70% expansion in AI use cases since Biden left office is staggering on its own terms. But the specific uses are what should be keeping people up at night. Pre-crime prison classification. Crisis line surveillance. Ideological filtering of children's services funding. These are not science fiction premises. They are line items in a government spreadsheet that almost no one will ever read, described in one sentence each, with no public comment, no external review, and no meaningful accountability structure attached to them.
The government is automating decisions about your freedom, your health, and in some cases your life. It is doing this at scale, at speed, and with the bare minimum of public acknowledgment it can legally get away with. The inventory is not a transparency document. It is a legal shield, something the administration can point to if anyone asks, proof that they told you, technically. They did not tell you. They filed a form.
