Somewhere inside the Department of Homeland Security, an AI system is scanning biometric data at the border. At the Department of Health and Human Services, another is helping process health program decisions that touch millions of Americans. Across the hall of nearly every federal agency, similar tools are quietly at work. The total count, according to a Government Accountability Office report published earlier this year: 3,611 individual AI use cases, a 69% increase over 2024 and roughly five times the number agencies reported in 2023.
The acceleration is striking. But the number alone does not answer the question that matters most to taxpayers, civil liberties groups, and the people whose benefits, travel, and legal cases now pass through automated systems: Is anyone making sure all of this actually works?
How the count grew so fast
The surge traces back to a series of federal mandates that forced agencies to inventory their AI systems and make those lists public. Executive Order 13960, signed in December 2020, first required the cataloging effort. The CIO Council updated its reporting guidance in 2021, 2023, and 2024, each round tightening the definition of what counts as an AI use case and expanding the agencies required to disclose. Office of Management and Budget memos, including M-25-21, now require ongoing reporting.
That means part of the five-fold jump since 2023 reflects genuine growth in AI adoption, and part reflects better bookkeeping. A system that was running in 2022 but never inventoried shows up as “new” when an agency finally reports it. The GAO’s methodology accounts for some of these shifts, but the public data does not cleanly separate real expansion from improved disclosure.
The policy landscape has also shifted. The Biden administration’s Executive Order 14110, issued in October 2023, imposed additional AI safety and transparency requirements on federal agencies. In February 2025, the Trump administration revoked that order and issued its own directive emphasizing AI adoption and reduced regulatory barriers. The current inventory numbers reflect systems deployed under both policy regimes, and the competing priorities of safety-first oversight versus rapid deployment create tension that agencies are still navigating.
Where the AI is concentrated
Several departments account for an outsized share of the 3,611 total. HHS maintains a dedicated inventory listing applications tied to health research, benefits administration, and public health surveillance. DHS publishes its own inventory files in spreadsheet format, covering biometrics, border operations, and cybersecurity. The Department of Justice reported 315 entries in its 2025 inventory, a 30.7% increase from the prior year, according to the DOJ’s AI use case inventory page.
The mandate reaches well beyond Cabinet-level departments. Even the Federal Reserve Board has published an AI use case inventory under the AI in Government Act and OMB guidance. That breadth helps explain the rapid growth: agencies that previously reported nothing, or listed only a handful of tools, are now disclosing dozens or hundreds of applications as standards tighten.
Agencies have also begun formalizing internal governance. The GAO report describes the emergence of chief AI officer roles, cross-functional review boards, and AI working groups inside departments. These structures vary in authority and maturity, but their existence signals that federal leaders recognize AI as a mainstream operational technology, not a side experiment.
What the inventories do not reveal
The 3,611 figure is a useful headline, but it obscures critical details. The GAO’s public-facing materials do not break the total down by risk category. There is no easy way to determine how many of those cases involve high-stakes decisions affecting individuals, such as benefits eligibility screening, criminal risk scoring, or immigration adjudication, versus routine automation like document summarization or IT helpdesk chatbots.
Generative AI applications are particularly hard to track. The GAO report addresses generative AI management practices but does not provide granular case studies. The DHS inventory, for example, names broad categories like biometrics and cyber defense without always specifying whether a tool is a generative model or a traditional machine learning classifier. That distinction matters: generative systems carry distinct risks around hallucination, accuracy, and misuse that current reporting templates may not fully capture.
Performance data is largely absent. Public inventories rarely include error rates, bias assessments, or the results of impact evaluations. Without that information, outside experts cannot judge whether a given AI tool is improving outcomes, merely shifting workloads, or introducing new risks. The inventories describe what exists. They do not describe how well it works.
Direct statements from agency officials about implementation challenges are also missing. The inventories are data tables, not narratives. Neither HHS nor DHS includes attributable quotes from leaders describing workforce training gaps, procurement bottlenecks, or difficulties managing vendor relationships for AI tools. The GAO provides independent analysis but does not extensively quote agency heads on operational friction, leaving a gap between policy commitments and the realities of deploying complex technology inside sprawling bureaucracies.
Why the stakes keep rising
The concentration of AI use cases in agencies like HHS and DHS carries direct consequences for the people those agencies serve. When HHS deploys AI in health programs, the accuracy and fairness of those tools can affect patient care and benefits processing for millions. When DHS applies AI to biometrics and border screening, errors or biases can affect travelers, asylum seekers, and immigration applicants. The inventories confirm these agencies are among the largest AI operators in the federal government, but they do not tell the public whether independent ethical audits are being conducted, how often systems are reviewed for disparate impact, or what remedies exist when automated decisions go wrong.
The political environment adds another layer of uncertainty. With the Trump administration prioritizing rapid AI adoption and rolling back some of the Biden-era guardrails, oversight groups and congressional watchdogs face pressure to ensure that speed does not come at the expense of accountability. The GAO’s continued auditing role is one check on that process, but the watchdog’s recommendations are non-binding, and agencies are not required to act on them.
For now, the documented 3,611 use cases confirm that AI is embedded across the federal landscape, from scientific research labs to law enforcement databases. Agencies are under legal obligation to disclose these systems, and oversight bodies are scrutinizing them more closely than at any point in the technology’s history inside government. But until those inventories are paired with meaningful data on performance, risk, and accountability, the public record will show how much AI the government is using. It will not yet show whether that use is consistently safe, lawful, or serving the people it is supposed to protect.
More from Morning Overview
*This article was researched with the help of AI, with human editors creating the final content.
