Medicare's WISeR AI prior authorization: denials were issued, but no record that the judgment was independently verified against the patient's own evidence

TL;DR

Since January 2026, US Medicare has piloted WISeR, an AI-assisted prior-authorization model, across six states. On June 23, 2026, KFF Health News (via CBS) reported that doctors attribute some denials to AI hallucinations garbling clinical facts — a patient needing a neck injection denied because “the thoracic region” was ineligible; a patient whose records stated four times that there was no numbness, denied on the grounds of numbness. CMS requires a licensed clinician, not a machine, to make the final non-coverage call. The adjudicating layer was there. What was missing is the layer that independently verifies, before a denial is issued, that the determination rests on the patient’s own evidence.

What happened

• The subject is WISeR (Wasteful and Inappropriate Service Reduction), introduced into fee-for-service (Original) Medicare. It covers 13 services deemed vulnerable to fraud or misuse — skin and tissue substitutes, electrical nerve stimulator implants, knee arthroscopy for osteoarthritis, and kyphoplasty for spinal fractures — across Arizona, New Jersey, Ohio, Oklahoma, Texas, and Washington. Inpatient-only services, emergency services, and services for which delay would pose a substantial risk are excluded.
• Doctors and their staff say some denials look like AI hallucinations that garble or invent clinical information. Earlier this year, nearly 100 patients at the University of Washington’s medical system alone were waiting for epidural injections because of WISeR-related delays, according to an April report from a US senator’s office drawing on hospital association data.
• Per CMS’s own announcement, model participants are paid based on their ability to reduce inappropriate utilization and lower Original Medicare spending, adjusted for quality and process measures covering decision speed and stakeholder experience. One vendor — the CEO of Humata Health — says he is not aware of any such hallucinations.

A denial reaches the patient through the following chain.

1. A physician submits the supporting medical records to an online portal ahead of one of the 13 covered procedures.
2. AI and machine learning grant an “immediate yes” to requests meeting the program’s criteria — one vendor puts this at about 88% of cases where the clinical data supports approval.
3. The rest go to review, and some come back denied on garbled clinical facts.
4. The patient faces delay, extra visits, or giving up on treatment; appeals rise, and review costs accrue on the government’s side as well.

Timeline — disclosure and response

• 2025-06-27: CMS announces the WISeR model, stating explicitly that final determinations that a request does not meet coverage requirements will be made by licensed clinicians, not machines.
• 2026-01 (mid-month): the pilot launches in six states. The outgoing CEO of the Ohio State Medical Association calls the rollout “quicker than normal” for the federal government; the policy director of the Washington State Medical Association says doctors “just sort of had to figure it out.” A vendor concedes an “aggressive rollout from the time of being notified to going live.”
• 2026-04: Sen. Maria Cantwell’s office publishes a snapshot report on WISeR drawing on hospital association data, including the roughly 100 patients waiting for epidural injections at the University of Washington.
• 2026-06-23: KFF Health News (Darius Tahir, published by CBS News) reports the confusion, errors, and delays, along with denials that doctors suspect originated in AI hallucinations.

This Brief does not adjudicate whether any individual case was an AI hallucination. CMS and the vendors state that humans make the final determination, and one vendor says he is not aware of any hallucinations. Individual physicians’ accounts are treated as accounts from parties with a direct stake. The senator’s office report could not be retrieved directly at the time of writing, so its figures are cited via KFF Health News’s reporting.

The response and industry movement after disclosure:

• CMS says the goal is “to reduce inappropriate care without delaying appropriate care,” with decisions returned within 72 hours. Clean claims are supposed to be paid within 15 days, according to physicians on the ground, but “six- to eight-week delays” (a Tulsa radiologist) are what they report seeing. At an April briefing, the vendor holding the Arizona contract acknowledged a large payment backlog stretching back to January.
• Appeals are reported to be rising, and CMS said it had factored in changes in the volume of appeals and the associated costs. It says there are “currently no changes” contemplated for the list of covered services, but that it continues to assess whether changes are warranted.
• The director of the CMS Innovation Center also acknowledged that “the percentage of providers committing waste, fraud, and abuse is small.”

Why it wasn’t stopped

The failure here is neither that an AI can be wrong nor that human review was absent from the design. There was no layer that independently verified and authorized the determination against the patient’s own evidence before the denial was issued.

Human final judgment is written into the design. CMS requires a licensed clinician to make the non-coverage determination, and vendors describe it the same way. On paper, the verifying layer existed. What was missing came earlier — the verification that reconciles a determination against its actual evidence before the action becomes final. When a physician states the same fact four times in the record and the misreading is not corrected, that layer is not operating effectively ahead of the action.

A denial is an action. Without a record independently confirming, before that action, that the determination rests on the right patient’s own record, the error takes effect first — as an adverse action already imposed.

The payment design can widen this gap. When payment is tied to reduced utilization and lower spending, the legitimacy of the determination has to be carved out as an independent record, or there is no way to show, after the fact, that the incentive did not tilt the determination. This descends directly from Brief 078, where erroneous automated eligibility determinations cut off Medicaid coverage, and it shares the structure of public-sector AI judgments that go unverified before action. The issue is not the “fairness” of the determination but the single point shown in Brief 012 (facial recognition and a wrongful arrest): that a determination is not independently verified and authorized before a coercive action follows.

What proof would have changed

Proof-as-auth inserts one layer into the path ahead of each act of finalizing an adverse determination: an independent verification that the determination rests on the underlying evidence. Rather than treating human final review as a stated policy that stands in for verification, it establishes — before the action can become final — whether the determination was applied to the right patient, against the right evidence, according to the stated criteria.

Lemma’s design against this primitive:

• A record before the determination is acted on. Before an adverse action is finalized, record that the determination was independently verified and authorized against the underlying evidence.
• Bind the determination to the evidence. Tie it to the patient record actually consulted, making misreadings — the wrong anatomical region, findings that contradict the file — detectable before the action.
• Provenance of the decision. Record the path from the AI’s recommendation to the human’s final call in a form that cannot be altered afterward, so that “a clinician decided” survives as a verifiable fact rather than a stated policy.
• Separate from the incentive. Even where payment tracks reductions, carve out the record of a determination’s legitimacy as an independently verifiable object.

Lemma is not a product that determines medical necessity, nor one that detects AI errors. Its scope is to verify independently, before an adverse action becomes final, that the determination rests on the underlying evidence, and to hold actions that carry no such proof. Accuracy work and human final review (clinician review, appeals, after-the-fact correction), on one side, and pre-execution proof (a record reconciling determination against evidence before the action), on the other, are complementary, not alternatives. The first corrects errors afterward; the second closes the space between “a determination was issued” and “a determination was independently verified and authorized against the evidence” — the one place detection structurally cannot reach. For the complementarity framing see “The last layer left for cyber defense in the age of AI” (Lemma, 2026-05); for design detail, “Proof-as-Auth: sign in without ever sending your key”; for scope, Pillar 02 — Verifiable AI.

Sources

• KFF Health News / CBS News (independent reporting): Darius Tahir, “Medicare’s AI push snarls patients and doctors in errors and delays” (2026-06-23) — https://www.cbsnews.com/news/medicare-ai-program-wiser-prior-authorization-errors-delays/
• CMS (primary, announcement): “CMS Launches New Model to Target Wasteful, Inappropriate Services in Original Medicare” (2025-06-27) — https://www.cms.gov/newsroom/press-releases/cms-launches-new-model-target-wasteful-inappropriate-services-original-medicare
• CMS (primary, operational guide): “WISeR Model Provider and Supplier Operational Guide” — https://www.cms.gov/priorities/innovation/files/wiser-provider-supplier-guide.pdf