Every major health insurance company in America made the same pitch to Wall Street this year. The message was simple: artificial intelligence will help us make faster coverage decisions and save money.

They left out the part where "faster" means denying your claim before a human ever sees it.

New research from Stanford University has landed like a grenade in the health insurance industry. The findings are devastating in their simplicity: when you train an AI system on historical claims data from a system already plagued by wrongful denials, the AI does not fix the problem. It scales it. The algorithm learns to deny care at industrial speed, replicating every bad decision that came before it, but now doing it millions of times per day instead of thousands.

"There is a world in which using AI could make that worse, or at least replicate a bad human system, because the data that it would be training on is from that bad human system," said Michelle Mello, a co-author of the Stanford study.

The Numbers Tell the Story

Class action lawsuits have already been filed against multiple insurers, accusing them of using AI to wrongfully withhold treatment. The cases allege a pattern: patients submit claims, AI systems reject them in seconds without meaningful review, and the appeals process is designed to exhaust people into giving up. This is not a bug. For the insurers, this is the feature they sold to Wall Street.

The Trump administration is not sitting this one out. It is actively testing AI for prior authorization decisions in the Medicare program, exploring whether artificial intelligence can speed up the process of approving or denying coverage for America's most vulnerable patients. At the same time, the White House is pushing to override state-level AI regulations that could limit how insurers use these systems.

Let that sink in. Red and blue states alike have been racing to limit AI in insurance decisions. The federal response? Remove the guardrails.

Why This Is Different From Every Other AI Story

Most AI stories are about theoretical risk. This one is about people not getting medical treatment they need because a machine said no. The Stanford researchers found real positives alongside the risks, but the core problem is structural: the training data is poisoned by decades of a system that incentivizes denial.

States like Alabama are already moving. SB 63 would regulate AI use in health coverage determinations. Georgia is debating SB 444, which would prohibit coverage decisions based solely on AI systems. But with the White House actively trying to preempt state regulation, these protections may not survive.

Here is the uncomfortable truth: health insurers have found the perfect use case for AI. Not to help patients, but to deny them faster, cheaper, and at a scale no human claims department could ever match. And the people who could stop it are too busy arguing about whether AI chatbots are too woke.

First reported by KFF Health News, with research published by Stanford University.