On April 14, the Trump administration disclosed that 3,611 active or planned AI systems are now operating across federal agencies—a 70% increase from the Biden administration’s final inventory—with virtually no public announcement or meaningful oversight.
The scale of this deployment matters because it represents a wholesale transfer of consequential decisions from human judgment to algorithmic systems: who receives federal grants, which prisoners get locked in solitary confinement before committing infractions, which veterans in crisis get routed to automated assessment, and how nuclear reactors respond to safety incidents. The Office of Management and Budget published the inventory on GitHub, a platform most Americans never check, ensuring that the deployment would pass largely unnoticed.
- The Scale of Deployment: 3,611 AI systems are now active or planned across federal agencies—a 70% increase from the previous administration’s inventory—disclosed with no formal public announcement.
- The Decisions at Stake: These systems determine federal grant eligibility, prisoner confinement conditions, veteran mental health assessments, and autonomous nuclear reactor responses, with minimal human review.
- The Transparency Gap: Most deployments are classified as non-“high impact” use cases, exempting them from transparency requirements and leaving citizens with no formal mechanism to contest algorithmic decisions that affect them directly.
The specific use cases reveal the depth of the shift. The Department of Health and Human Services hired Palantir—a company with a documented history of work for the military, CIA, and ICE—to scan all grant applications and flag those deemed “not ideologically aligned” with administration dictates. The Federal Bureau of Prisons is deploying an AI system to assess “potential for misconduct for newly admitted inmates,” routing people into high-security confinement based on algorithmic prediction rather than actual conduct. The Department of Veterans Affairs is building a system that will listen to calls on the veterans crisis line, then cross-reference external databases to assess mental state and suicide risk.
What distinguishes this moment from routine government technology adoption is the absence of disclosure infrastructure. Unlike the Department of Justice system, which theoretically involves public consultation, most of these deployments are classified as non-“high impact” use cases—a label applied inconsistently across agencies—and therefore exempt from transparency requirements. Unless you follow FedScoop or monitor the OMB’s GitHub account, you would have no way of knowing these systems exist.
• A July 2025 GAO review of 11 selected federal agencies found that the total number of reported AI use cases has grown sharply, with inventory reporting requirements inconsistently applied across departments.
• 3,611 AI systems now active or planned across the federal government—up 70% from the prior administration’s count.
• The majority of use-case descriptions in the OMB inventory span a single sentence, with no mandatory disclosure of error rates, training data, or safeguards.
How Does the Federal AI Deployment Mirror Cambridge Analytica’s Architecture?
This pattern echoes a structural problem that defined the Cambridge Analytica scandal: the harvesting of behavioral data at scale without meaningful consent, followed by the application of that data to influence individual outcomes. In CA’s case, the firm vacuumed up psychological profiles of millions of Americans via Facebook, then weaponized that data to micro-target voters with personalized messaging designed to suppress turnout or shift preference. The consent was buried in terms of service; the targeting was invisible to those affected.
Here, the mechanism differs—government agencies are the data collectors and decision-makers rather than private firms—but the architecture is structurally identical: mass data collection, algorithmic inference about individual behavior and risk, and consequential decisions made without the knowledge or input of those affected. A veteran calling a crisis line does not consent to having their mental state assessed by an algorithm trained on external databases. A grant applicant does not know their application will be scanned for ideological alignment. A newly incarcerated person has no opportunity to contest the algorithmic prediction that will determine their confinement conditions. As documented in analyses of Cambridge Analytica’s legacy, the most dangerous feature of these systems is not any single decision they make—it is the invisibility of the process to those most affected by it.
• GAO’s AI accountability framework identifies continuous monitoring, transparent deployment documentation, and defined human oversight as the minimum requirements for responsible government AI use—standards the current federal inventory largely does not meet.
• The framework specifically flags the risk of deploying AI in high-stakes administrative decisions without validated impact assessments or public accountability mechanisms.
• Where those safeguards are absent, the framework warns, errors in algorithmic classification become structurally invisible and practically uncontestable by those affected.
What Does the Inventory Actually Disclose—and What Does It Hide?
The inventory itself reveals how thin the disclosure actually is. Most use-case descriptions span a single sentence, rarely exceeding a paragraph. There is no mandatory explanation of how the AI works, what data it uses, what error rates have been documented, or what safeguards exist if the system makes a mistake. The Department of Energy’s plan to use AI to autonomously control nuclear reactors in response to safety incidents is described with minimal technical detail. The VA’s crisis-line listening system includes no information about how the algorithm was validated or what happens when it misclassifies risk.
The source material notes that some of these applications may have legitimate purposes. Using statistical methods to assign prisoner security classifications predates AI by decades, though research suggests such systems are often biased and ineffective. Autonomous model predictive control of nuclear reactors is a studied, widely applied aspect of plant management. Machine translation systems deployed by Customs and Border Protection could improve communication when human interpreters are unavailable. The concern, then, is not necessarily that AI is being used in government—it is that the deployment is happening without the transparency, validation, or public deliberation required to distinguish beneficial applications from harmful ones.
Why Are Other Democracies Setting a Higher Bar?
Other democracies have established significantly higher standards. Canada launched an AI registry in 2025 paired with a federal directive requiring transparent risk-scoring and impact assessments for systems that make administrative decisions about citizens. France’s 2016 Digital Republic Act mandates that all algorithms automating government decisions be subject to public records requests, appealable to human reviewers, and disclosed to those affected. Washington DC and California are actively conducting public deliberations on where and how AI should be used in government.
The US federal government has adopted neither approach. The inventory exists, but it functions as disclosure theater—technically public yet practically invisible, containing minimal information and triggering no mandatory public comment period. A citizen would need to know the system exists, navigate to an obscure GitHub repository, parse technical jargon, and then have no formal mechanism to contest or challenge what they find. This dynamic—where the architecture of manipulation is technically visible but practically inaccessible—is precisely what scholars examining the Cambridge Analytica era identified as the defining feature of systems designed to evade accountability without technically concealing themselves.
What Happens Next—and Who Is Watching?
The question now is whether this deployment pattern will persist or whether pressure for transparency will force the administration to adopt the kind of algorithmic impact assessment and public deliberation processes that other democracies have established. The broader context matters here: the same logic that allowed surveillance capitalism to normalize mass behavioral profiling in the private sector—moving fast, classifying data use as proprietary, and burying consent in fine print—is now being replicated inside the institutions that hold the most coercive power over citizens’ lives.
That answer will likely emerge over the coming months as civil society organizations and privacy advocates begin examining the full inventory. The precedent being set now—3,611 systems, minimal disclosure, no mandatory public comment—will define the baseline against which any future accountability framework must fight. Whether that baseline shifts depends on whether the public, the press, and oversight bodies treat a GitHub repository as the transparency it is claimed to be, or the theater it actually represents.
