The Algorithm Said No

Automated processing systems are slashing Medicare coverage — and violating due process

Just before Christmas in 2022, Robert Austin and his young daughter were living in his car in El Paso. He couldn’t find a shelter that would take them both, he turned to the safety net: he applied for food stamps, emergency cash assistance, and Medicaid for his girl.

They qualified, but they were denied. Each time Austin called to ask why, he found himself stuck in the same loop, unable to reach a human who could explain — or take responsibility for — the system’s decision. Lawyers at Texas RioGrande Legal Aid later identified the culprit: the state’s automated eligibility system, built by Deloitte, had been wrongly denying people by the thousands. A father and his daughter spent the holidays in a car because a program said no, and no one could say why.

This is the new way the government says no. Quietly, without a vote, agencies have handed hard decisions to software and left people with no real way to appeal. This violates the oldest promise in our law. Due process means that before the state takes your liberty or your property, it owes you a reason, a chance to respond, and a human being obligated to listen. That promise exists because the government is powerful — and often wrong. A right you cannot contest is not one you truly hold.

Automated systems strip away all three guarantees of due process at once. They replace the reason for a score no one can fully explain, the hearing for an appeal against a black box, and the human decision-maker for a clerk who rubber-stamps the machine. They also flip the burden, so that you, not the state, must disprove the verdict.

None of this is new, and we were warned. In 2009, Indiana’s automated system cut off a cancer patient, Omega Young, for missing a recertification appointment because she was in the hospital. She won her appeal the day after she died. Michigan’s MiDAS accused tens of thousands of unemployment fraud; an audit found 93 percent of its flags were wrong, and the state paid $20 million. Arkansas let an algorithm slash disabled residents’ home care and was struck down in court. Each was caught years late, through litigation, after the damage was done.

So how did this happen? Vendors sell the systems as fraud control; agencies buy them under budget pressure and skimp on the safeguards. Automation is cheap, and a machine’s verdict looks neutral, so a wrong number feels more defensible than a caseworker’s judgment. When one breaks, no one answers for it, because “the system decided” sounds like nobody’s fault, and the logic is usually a trade secret no one can inspect.

What is new is that algorithmic bias is becoming more often a design feature, not a bug, that profits by saying no, and it has moved from the margins into Medicare. On January 1, 2026, Medicare switched on an AI program called WISeR that screens certain doctor-ordered services through prior authorization and pre-payment review in six states, with contractors paid from the savings their denials produce. Austin’s runaround is now built into Medicare. And in December, the White House ordered the Justice Department to sue states that try to regulate AI, even as Washington deploys it. The states were the only layer catching the errors. Stripping them while expanding the machines is not negligence. It is a choice.

The strange part is that the law protecting us already exists. Goldberg v. Kelly (1970) requires the government to give notice and an evidentiary hearing before it cuts off benefits to welfare recipients. Additionally, the 1988 Computer Matching and Privacy Protection Act bars cutting anyone off based on a computer match without first verifying it and letting them respond.

So how are federal and state agencies getting away with it? By keeping the form and skipping the substance. The notice arrives, but its “reason” is a code you can’t contest; the appeal exists, but only after the harm is done; and the logic stays sealed as a trade secret, so the right to refute becomes the right to argue with a locked box. And the rules bite only when someone sues. Kevin De Liban, the legal-aid lawyer who beat Arkansas’s care-cutting algorithm, now runs a nonprofit built for this fight, but one lawyer cannot keep pace with a nationwide rollout. Enforcement depends on a lawsuit, and the people these systems hit hardest, the poor, the sick, and the disabled, are the least able to bring one. The harm falls, by design, on those who can’t fight back.

The federal protections are sound on paper but almost never enforced. And now Washington is in court to kill even the state rules that were starting to make them stick. In April, the Justice Department’s new AI task force joined an Elon Musk company in suing to strike down Colorado’s first-in-the-nation rule guaranteeing notice, explanation, and human review of automated decisions in certain sectors. No court has ruled. The threat alone, backed by the federal dollars a state cannot afford to lose, was enough: within weeks Colorado gutted its own law.

Due process cannot be delegated. A machine can recommend, but it should never decide. Every denial must carry a real reason a person can contest, with its logic open to appeal, with benefits intact until that appeal is heard. Congress should set these protections as a national minimum that states can strengthen, never as a cap that overrides them. Behind every no there must be a human who can be made to explain it, and answer when it is wrong. Otherwise, we have a government that can refuse you anything and explain nothing. That is not a government at all. It is a vending machine that keeps your money.

---

Reuben Steiger is a writer and entrepreneur based in Princeton, NJ. Over a 25-year career he has helped start companies including Second Life and has led global innovation for companies including Interpublic and Omnicom. His current focus is the scaling and adoption of AI technologies. He collects books about the future.