The method itself is an incentive problem

Prior authorization — an insurer deciding whether a treatment counts as 'necessary' — is, used judiciously, a check on unnecessary spending. But a 2025 AMA survey found 61% of physicians worry AI tools will worsen denials of care they consider necessary, and a new Commonwealth Fund survey found one in five privately insured Americans experienced a denial in 2025; 41% of those said it delayed care, more than a quarter said their condition worsened. CMS's WISeR model, running through December 2031, pairs machine learning with human clinical review to screen services like knee arthroscopy.[1]

This time I'm not questioning scientists' honesty but the program designers': the vendors running the model earn a share of 'averted expenditures.' So a denial, right or wrong, is directly a revenue line. That's the inverse of 'not a miracle, a method' — the method itself rewards the profitable outcome, not the correct one. A 2022 HHS Office of Inspector General report found Medicare Advantage plans had denied more than one in ten requests that appeared to meet coverage rules; 81% of appeals succeeded in 2024 — but appealing takes time and know-how many patients don't have.[1]

Two directions, one administration

What's striking is that the same administration is expanding AI-assisted prior authorization in original Medicare while telling private insurers to 'reduce' it. CMS Administrator Mehmet Oz told insurers: "If you don't do it yourselves, then we're going to do it for you." The industry's own survey shows prior-authorization requests fell 11% between June 2025 and April 2026 — but whether the denial rate fell is unknown. The count can drop while the ratio stays the same.[1]

My guess: WISeR's first-year data across the six pilot states — including the delay and denial cases already flagged by the Washington Post, KFF Health News and the Seattle Times — turns into a congressional inquiry or funding-block legislation within months. If a billion years of evolution has taught us anything, it's that no algorithm makes a system 'intelligent' until you've fixed the incentive running underneath it.[1]