AI coverage decisions draw scrutiny

A doctor can say a scan, rehab stay, or cancer drug is medically necessary, and the insurer can still stop the bill until it clears a checkpoint called prior authorization. Now that checkpoint is increasingly being run with artificial intelligence tools that insurers say will cut costs. (kffhealthnews.org) KFF Health News reported on April 10 that executives at nearly every major health insurer told Wall Street this year that artificial intelligence would help make coverage decisions more cheaply. The same report said the Trump administration is also testing artificial intelligence in Medicare’s prior authorization process. (kffhealthnews.org) The pitch is simple: a human nurse or doctor reviewer is slow and expensive, while software can sort thousands of requests like an airport scanner sorting bags. Stanford researchers wrote in February 2026 that insurers and providers are rapidly adopting these tools to process prior authorization requests and claims. (hai.stanford.edu) The problem is that software learns from old decisions, and old insurance decisions already include delays and wrongful denials. Stanford’s Michelle Mello warned that training artificial intelligence on a “bad human system” can replicate or worsen the same mistakes at machine speed. (kffhealthnews.org, hai.stanford.edu) That speed changes the economics of denial. Stanford said prior authorization was already “fraught” before artificial intelligence, and making reviews cheaper could “supercharge” a process that already blocks care. (hai.stanford.edu) Doctors say they are already seeing the effect. In an American Medical Association survey released on February 24, 2025, 61% of physicians said they were concerned insurers’ use of artificial intelligence was increasing prior authorization denials. (ama-assn.org) The same American Medical Association survey found 29% of physicians said prior authorization had led to a serious adverse event for a patient in their care, including hospitalization, permanent impairment, or death. When a system with those stakes gets automated, a bad rule can spread the way a typo spreads in a copied spreadsheet. (ama-assn.org) Courts are now being asked whether some insurers crossed that line already. A federal judge in February 2025 allowed key parts of a class action against UnitedHealth to continue after plaintiffs alleged an algorithm was used to deny post-acute care in Medicare Advantage plans instead of individualized medical review. (healthcarefinancenews.com, statnews.com) A separate case against Cigna also survived in part, with plaintiffs alleging its PxDx software denied claims without examining each case the way California law requires. Those lawsuits do not prove every automated system is unlawful, but they show the fight has moved from conference rooms to courtrooms. (courthousenews.com, nfp.com) Federal regulators are trying to draw a line, but it is a narrow one. The Centers for Medicare & Medicaid Services said in February 2024 that Medicare Advantage plans may use artificial intelligence in coverage determinations, but decisions still must be based on the individual patient’s circumstances and must not produce biased or discriminatory results. (reedsmith.com, cms.gov) That line is getting harder to police as Medicare itself experiments with automation. The Electronic Frontier Foundation sued the Centers for Medicare & Medicaid Services on March 25, 2026, seeking records about a multi-state model using artificial intelligence to evaluate requests for medical care, including documents on accuracy, bias, and hallucinations. (eff.org, healthcaredive.com) The argument over artificial intelligence in health coverage is not really about whether software can read forms faster than people. It is about whether a tool built to save insurers money can be trusted at the exact moment a patient is trying to get a surgery date, a rehab bed, or a drug refill approved. (kffhealthnews.org, hai.stanford.edu)