The Trump administration’s OMB inventory reveals a sweeping expansion of AI across sensitive government functions, but transparency and oversight mechanisms remain thin.
Summary
- The US Office of Management and Budget has disclosed 3,611 active or planned federal AI use cases — a 70% increase on the previous administration’s inventory.
- Applications span high-stakes domains including nuclear reactor oversight, veteran mental health crisis lines, inmate misconduct prediction, and grant application screening.
- Most disclosures amount to a single sentence with no public consultation requirement, and the ‘high impact’ classification that triggers oversight is applied inconsistently.
- International comparators — France’s Digital Republic Act and Canada’s algorithmic impact assessment directive — offer concrete models for more rigorous accountability.
- The core risk for organisations and the public is not AI in government per se, but AI in government without adequate transparency, appeal mechanisms, or independent validation.
Scale of deployment
On 14 April, the US Office of Management and Budget quietly released its 2025 inventory of federal AI use cases, tallying 3,611 active or planned applications across agencies. That figure represents a 70 per cent increase on the final inventory published under the Biden administration. The breadth of the list is striking: AI is being inserted into decisions touching individual liberty, public health, nuclear safety, and border operations.
What is actually on the list
A handful of cases illustrate the range. The Department of Veterans Affairs is developing an AI system to monitor calls to the veterans crisis line and draw on external databases to assess suicide risk. The Federal Bureau of Prisons is building a tool to predict misconduct risk for newly admitted inmates — before any incident has occurred in custody — and route individuals to higher-security conditions accordingly. The Department of Energy is testing AI for autonomous response to nuclear safety incidents. Meanwhile, the State Department has ended a program that used AI to forecast mass civilian atrocities, which had been designed to support conflict prevention.
Context matters, and is largely absent
Bruce Schneier and co-author Nathan E. Sanders, writing in The Guardian, note that alarming-sounding applications are not automatically unjustified. Predictive methods in prisoner classification have existed for decades, even if those systems have often been criticised for bias and limited effectiveness. Autonomous model predictive control in nuclear plant management is a well-studied and widely applied discipline, and the AI component now disclosed was reportedly initiated under the previous administration. The concern is less with the concept of AI in government and more with what is missing from the disclosures: purpose, methodology, validation, and accountability.
Transparency in name only
The inventory entries are typically a single sentence and rarely exceed a paragraph. Public consultation is not mandated for the majority of use cases because most are not classified as ‘high impact’ — a designation applied inconsistently across agencies. Of the specific examples cited in the source material, only one agency even proposed involving the public. The disclosures are buried in a GitHub repository maintained by the federal chief information officer’s office, and were largely unreported outside specialist outlets such as FedScoop.
Where better practice exists
The authors point to France’s 2016 Digital Republic Act as an international benchmark. The law requires that all algorithms used to automate government administrative decisions be subject to public records requests, be appealable to a human reviewer, and trigger mandatory notification to anyone affected. Canada’s 2025 AI use case registry is accompanied by a longstanding federal directive requiring transparent risk-scoring, impact assessments, and stakeholder consultation from the design stage of any automated decision system affecting citizens. Both examples demonstrate that disclosure can be meaningful rather than perfunctory.
Not all AI in government is problematic
The source material is careful to avoid a blanket critique. Machine translation is offered as a practical example: Customs and Border Protection has deployed AI translation to assist officers when human interpreters are unavailable. The 2025 inventory includes 70 such translation use cases, up from 58 in 2024. The authors’ position is that AI can improve the efficacy, efficiency, and accessibility of government services — but that public trust depends on genuine transparency and the ability to contest automated decisions.
Why it matters
For security executives, the US federal AI inventory is a signal, not just a policy story. Organisations that supply technology, data, or services to government agencies — or that operate in regulated sectors — should expect AI-driven decision-making to become a standard feature of their regulatory and compliance environment. The inventory also surfaces a governance gap that mirrors challenges inside private enterprises: AI is being deployed faster than accountability frameworks can keep pace. CISOs should note that inadequate transparency, absent appeal mechanisms, and inconsistent risk classification are not unique to government; they are common failure modes in enterprise AI programmes as well. If federal agencies with significant resources cannot produce coherent impact assessments, the bar for private-sector AI governance is implicitly raised for those who wish to demonstrate better practice.
What to do now
- Audit your organisation’s AI use case register — if one exists — to verify that entries contain sufficient detail on purpose, methodology, data sources, and validation, rather than single-sentence descriptions.
- Assess whether AI systems that make or inform decisions affecting individuals have a defined appeal or human-review pathway, consistent with the standard set by frameworks such as France’s Digital Republic Act.
- Review how your organisation classifies ‘high impact’ AI use cases and whether that classification triggers proportionate governance controls, including stakeholder consultation.
- Monitor regulatory developments in jurisdictions where your organisation operates; Canada’s algorithmic impact assessment model and similar frameworks may inform future requirements.
- Where your organisation supplies technology or services to government agencies, understand whether the agency’s AI governance obligations extend to your systems or data through contractual or regulatory mechanisms.
