Political Tech
Federal agencies have spent years writing AI policy documents — but translating those frameworks into actual governance is proving far more difficult than anyone anticipated.
NewsOnScale Staff
July 8, 2026
There is no shortage of AI policy in Washington. There are executive orders, agency memoranda, interagency task forces, and reams of published guidance. What there is a shortage of is execution — and that distinction is starting to matter enormously.
A new analysis from the Brookings Institution lands at a pivotal moment: federal agencies are being asked to move from the deliberative phase of AI governance into something far harder, embedding actual accountability structures into day-to-day operations, procurement decisions, and public-facing services. The piece doesn't bury the lede. The machinery of federal AI oversight, such as it exists, was built for discussion. It was not built for enforcement.
## The Governance Gap Is Not Abstract
To understand why this matters, consider what federal AI deployment actually looks like right now. Agencies are using algorithmic systems to inform benefits determinations, flag fraud, screen job applicants, and assist in law enforcement. These aren't hypothetical use cases. They're live, and in many cases they're operating under governance frameworks that are either unenforced, unverifiable, or both.
The Office of Management and Budget issued guidance in 2024 requiring agencies to designate Chief AI Officers and publish inventories of their AI use cases. That was a meaningful step. But designation is not accountability. Publishing a list of AI systems is not the same as auditing them. And the distance between those two things — between documentation and genuine oversight — is where real harms tend to accumulate.
Brookings points toward a structural problem that NewsOnScale has tracked from the platform and private-sector side as well: governance bodies in the AI space are consistently given the mandate to monitor without being given the authority, resources, or independence to act. The result is the appearance of oversight without its substance.
## Who Actually Has the Power Here?
Part of what makes this moment so consequential is that the question of who owns federal AI accountability remains genuinely unsettled. Is it OMB? The AI Safety Institute, which has already faced political turbulence? Individual agency inspectors general, most of whom lack the technical capacity to audit machine learning systems meaningfully? Congress, which has passed no major AI legislation?
The answer right now is: approximately no one, in any coordinated sense. Responsibility is distributed across agencies with different risk tolerances, different technical capacities, and different political pressures. That diffusion isn't an accident — it often reflects deliberate choices by agencies and their political overseers to preserve flexibility. But flexibility at the governance layer tends to become unaccountability at the operational layer.
## The Procurement Vector Deserves More Scrutiny
One dimension that warrants particular attention — and that tends to get lost in high-level governance conversations — is federal AI procurement. When an agency buys or contracts an AI system from a private vendor, what due diligence is required? What performance standards are contractually enforced? What happens when a vendor's system produces discriminatory or erroneous outputs at scale?
These questions don't have clean answers in the current federal framework. Vendors operate in a relatively permissive environment where the reputational and legal risk of AI failures in government contexts remains low. That is a market signal, and it shapes what gets built and sold.
## What Real Execution Looks Like
Moving from governance to execution — to borrow the framing — requires a few things that are currently missing: technical capacity inside agencies to actually evaluate AI systems, not just inventory them; independent audit mechanisms with real authority; and procurement standards with teeth. It also requires political will to treat AI accountability as a standing obligation rather than a communications exercise.
None of that is impossible. Some agencies are further along than others. But the gap between the policy layer and the operational layer in federal AI is not closing on its own. Without structural intervention, what Washington has built is an elaborate framework for saying the right things while the systems underneath continue to run largely unchecked.