That is the problem worth naming now, while it can still be fixed cheaply. We are assembling an autonomous bureaucracy from thousands of small pieces, and we have not yet written its constitution.

The founders did not trust good men to govern well. They trusted structure. Power was divided so that the branch that wrote a rule could not also apply it and judge it. Every decision had a place to be appealed. Every office had a jurisdiction, and every official had a name that could be attached to a mistake. The administrative state grew up outside that design, which is why conservatives spent decades dragging it back inside. The Supreme Court advanced that effort in 2024 by ending judicial deference to agency interpretations of law.

The automated bureaucracy has none of those features, because each piece is too small to seem to need them.

Consider separation of powers. In a great many automated chains, the same pipeline flags a case, scores it, decides it, and triggers the enforcement action, with a human somewhere in the loop who sees a recommendation rather than a choice. The rule-writer, the applier, and the judge have collapsed into one system.

Consider appeal. A citizen can appeal a decision to a human. But if the human’s only tool is to run the same workflow again, the appeal is theater. A real appeal requires a person with the authority and the information to overrule the machine, and most automated chains have not been designed with that person in mind.

Consider jurisdiction. Automated decisions increasingly run across agencies. An identity check in one system feeds an eligibility model in another, which feeds a payment decision in a third. When the outcome is wrong, no single agency owns the chain, and the citizen is passed from one office to the next, each of which correctly says the error did not originate with it.

And consider accountability. When a caseworker denies a benefit, there is a name on the letter. When a model does it, there is a vendor’s product, an agency’s dataset, and an OMB reporting category. There is no one to hold responsible, which means there is no one with an incentive to notice.

Michigan learned this a decade ago. From 2013 to 2015, the state ran its unemployment fraud detection on an automated system with no human review. It falsely accused tens of thousands of residents of fraud, garnished wages, and drove families into bankruptcy, and it ran for nearly two years before anyone with authority stopped it. An error rate above 90% was later documented. Australia repeated the mistake on a national scale with a scheme its own royal commission found in 2023 to be neither fair nor legal, and the government ended up refunding roughly 430,000 debts. Malice had nothing to do with it. The machines kept running because nobody had jurisdiction to stop them, nobody was accountable for their output, and the appeal path led back to the machine.

Here is the part that should concern this administration. A civil servant can be reassigned, disciplined, or, under the reforms now underway, removed. A workflow cannot be fired. It can only be switched off, and only by someone who has the authority and knows where the switch is. An administration that has worked hard to make the bureaucracy answerable again should not want to replace it with one that answers to no one.

And whatever gets built now will be inherited. The autonomous bureaucracy assembled under this White House will be operated by the next one, and the workflows will not care who won the election. A constitutionalist who would never accept an unaccountable human agency should be the last person to accept an unaccountable machine one, because the machine version will outlast every political appointee who approved it.

The fix does not require slowing down. It requires four rules, and they belong in statute rather than in an OMB memo that the next administration can rewrite.

First, every automated decision that has a legal or material effect on a person bears the name of the official who owns it. Not the vendor. Not the agency. A person.

Second, every such decision has a right of appeal to a human who holds the authority to overrule the system, not only to rerun it, and who can see the evidence the system used.

Third, every automated chain that crosses agency lines has a single owning agency with jurisdiction over the whole chain, so that a citizen caught in the middle has one door to knock on.

Fourth, automated authorities expire. A workflow that makes consequential decisions should require reauthorization on a fixed schedule, as a spending program does, so the default is that it stops unless someone with a name decides it should continue.

TRUMP’S AI MANDATE: DEFINING KILLER ROBOTS WON’T SAVE US — HARDCODING HUMAN CONTROL WILL

The founders wrote a constitution before they trusted the government it created. They were right, and the principle has not changed because the government now runs on software.

The Trump administration has the chance to write the rules for the machines before the machines write the rules for us. That project belongs to neither party, because it is older than both.

Burak Oktenli is an independent researcher based in Washington, D.C. He holds a bachelor’s degree in computer science and engineering from the University of South Florida and an MBA, and is completing a master’s of professional studies in applied intelligence at Georgetown University. His writing focuses on defense strategy, alliance burden-sharing, military technology, and the governance of autonomous systems.