The federal government's own inventory now lists 3,611 individual uses of artificial intelligence across the executive branch, 1,818 of them already deployed or in pilot. Of those, agencies themselves classify 445 as "high-impact," meaning they can shape a decision with legal or material consequences for a citizen. The Department of Veterans Affairs alone reports 215. Taken one at a time, each is modest: a model that flags a claim, a workflow that routes an application, a tool that drafts a determination for a human to sign. Taken together, as one essay in the Washington Examiner argues this week, they are quietly assembling an autonomous bureaucracy that no one has written a constitution for.
The point is sharper than the usual worry about robots and red tape. The American administrative state was built to be checked: the office that writes a rule is not supposed to be the one that applies it and judges it, every decision has somewhere to be appealed, and every official has a name that can be attached to a mistake. An automated decision chain collapses those separations. The same pipeline can flag a case, score it, decide it, and trigger enforcement, with a human in the loop who sees a recommendation rather than a real choice. Appeal to a person whose only tool is to run the same workflow again is, as the essay puts it, theatre.
This is not hypothetical, and the history is ugly. From 2013 to 2015 Michigan ran unemployment fraud detection on an automated system with no human review; it falsely accused tens of thousands of residents, garnished wages, and ran for nearly two years before anyone with the authority stopped it. A later review put the error rate above 90 percent. Australia's Robodebt scheme repeated the mistake at national scale, and a 2023 royal commission found it neither fair nor legal, with roughly 430,000 debts ultimately refunded. In each case malice was absent. The machines kept running because no one owned the chain, no one was accountable for the output, and the appeal path led back to the machine.
The politics here are pointed. This build-out accelerated after an April 2025 White House budget office directive told every agency to speed up AI adoption and clear away the barriers. An administration that came to office promising to make the bureaucracy answerable again, to shrink it and make it possible to fire, is on course to replace it with one that cannot be fired at all. A civil servant can be reassigned or removed; a workflow can only be switched off, and only by someone who has the authority and knows where the switch is. Whatever gets built now will be inherited by the next administration, and the workflows will not care who won the election.
The proposed fix is modest and mostly procedural: put a named official behind every consequential automated decision, guarantee an appeal to a human who can actually overrule the system rather than rerun it, give every cross-agency chain a single owning agency, and make automated authorities expire unless someone reauthorises them. What is telling is that these belong in statute, not in a budget-office memo the next administration can rewrite with a signature, the same soft-law-versus-hard-law tension the Federation of American Scientists traces across AI policy more broadly. Only 40 percent of Americans expect AI's impact on society to be positive over the next two decades, and two-thirds have little confidence the government can regulate it. The government is now discovering that the hardest system to regulate may be its own.