Opinion: The Government's Fraud Algorithm Just Cut Off Health Coverage for 760,000 People

The Trump administration announced on Sept. 22 that it is canceling roughly 315,000 Affordable Care Act enrollments, covering more than 760,000 people, after flagging them as fraudulent or otherwise ineligible. In a press release from the Centers for Medicare and Medicaid Services, Health and Human Services Secretary Robert F. Kennedy Jr. said the government expects to recover about $2.2 billion in subsidies, while CMS Administrator Dr. Mehmet Oz said the agency is also terminating more than 200 insurance agents and brokers and has opened a six-month freeze on new broker registrations. Another 419,000 enrollments, Vice President JD Vance said, remain under review.

Nobody disputes that fraud exists in the ACA marketplace. What should trouble anyone paying attention is how the government decided who counts as fraudulent, and how little recourse the people it flagged actually have.

A Program With a Real Fraud Problem

The administration's underlying complaint is not invented. CMS's own 2025 Marketplace Integrity and Affordability rule, finalized last year through ordinary notice-and-comment rulemaking and published in the Federal Register, projected that between 725,000 and 1.8 million people would lose marketplace coverage in 2026 as the government tightened verification. That rule followed years of documented abuse: a KFF analysis notes that CMS logged 183,553 complaints of unauthorized enrollments and 90,863 complaints of unauthorized plan-switching on HealthCare.gov between January and August of 2024 alone, and suspended 850 brokers over the same stretch. Commission-hungry brokers enrolling people without their consent, or shopping them into cheaper subsidized plans without permission, is a real and well-documented problem, and the agency's push to terminate brokers who submitted applications missing Social Security numbers is a defensible response to it.

So the administration has a case. The question is whether the process it built to act on that case can actually tell a defrauded taxpayer from a low-income parent who simply didn't open a text message.

An Algorithm, Not a Hearing

According to Dr. Oz, the roughly 760,000 people losing coverage were identified because they had not filed a medical claim, had not filled a prescription, and had not responded to outreach by phone, email or text. Vance described the underlying method as pattern recognition, telling reporters that "for program eligibility and diseligibility based on fraud, it's not the same standard of proof as like a criminal conviction." Neither CMS nor the White House has released the model's accuracy rate, its false-positive rate, or the data it was trained on.

That matters because the criteria as described do not actually distinguish fraud from ordinary life. A healthy person who hasn't needed a doctor this year files no claims. A gig worker who moves apartments or changes phone numbers misses a text. Under the government's own rules for the program, enrollees are supposed to estimate their income in advance and reconcile it against actual earnings at tax time — a design that assumes and tolerates honest error. Conflating "didn't respond within a window" with "fraudulently enrolled" collapses that distinction. Reporting from MPR News adds a further detail: the administration sent lists of roughly a million people to insurers and asked the companies to make contact, with coverage canceled automatically if there was no response within 30 days. That puts the burden of proving eligibility on people who may never have known they were being tested, relayed through a for-profit intermediary rather than delivered to them directly by the government making the decision.

The Appeal Almost Nobody Files

The administration's answer to all of this is that enrollees can appeal. HealthCare.gov maintains a formal appeals process for exactly this kind of eligibility dispute. But Dr. Oz has undercut his own safeguard by repeating, in his and Vance's telling, that "when we do these big cutoffs, almost no one comes." An appeals process that its own administrator expects almost nobody to use is not much of a check on error — it is a legal formality that shifts the burden of accuracy from the government to a population that, by definition, already failed to respond to earlier notices.

"In most of those cases, if you aren't fraudulent, you can appeal back to the agency, explain your situation, and the agency will make payment. But as Dr. Oz has said repeatedly, when we do these big cutoffs, almost no one comes."

There is an instructive, if imperfect, precedent for what happens when an algorithm with a high error rate meets a population that rarely appeals. A class-action lawsuit against UnitedHealth Group alleges the insurer's nH Predict tool generated coverage denials for elderly Medicare Advantage patients that were reversed more than 90 percent of the time on appeal — while only about 0.2 percent of affected patients ever appealed at all. That is a different program and a private company rather than the government, so the analogy is not exact. But the mechanism is the same: when the entity making a high-stakes coverage decision knows that almost no one will contest it, the incentive to tolerate a high error rate grows rather than shrinks. Cynthia Cox, who directs KFF's ACA program, put the concern plainly, telling reporters that while fraudulently enrolled people should lose coverage, "whether this was the appropriate process by which to identify fraudulent enrollees, and also whether all of them were indeed fraudulently enrolled" is a separate and unresolved question, since the removals bypassed the ordinary regulatory process. Former CMS official Ellen Montz made a related point: the announcement offered no detail on how individual cases were actually determined.

What Accountability Would Look Like

None of this requires abandoning the goal of rooting out broker fraud, which is real and costly. It requires the administration to treat coverage termination with the seriousness it deserves. That means CMS should publish the criteria and error rates of whatever model or rules engine flagged these 760,000 people, the way agencies are expected to document the basis for consequential automated decisions. It means direct, plain-language notice to enrollees from the government making the decision, not a 30-day pass-through via insurers who have their own incentives. It means a genuine pre-termination review for cases that turn on ambiguous signals like an unanswered text rather than clear identity fraud such as a missing Social Security number. And it means congressional oversight — hardly the exclusive province of one party — sufficient to determine, before the next round of cuts, how many of the roughly 419,000 people still under review will lose coverage for having done nothing wrong at all.

A government that can strip health coverage from three-quarters of a million people on the strength of a pattern-recognition model, while acknowledging that almost no one will contest the decision, has built a system optimized for throughput rather than accuracy. Fixing actual fraud and protecting people who did nothing wrong are not competing goals. But right now, only one of them is getting resourced.