New Records Obtained by EFF Expose More Problems with Medicare’s AI Experiment
EFF obtained 800+ pages of records on WISeR, a Medicare program that uses AI to screen prior authorization requests, and they show a rushed launch with delays, wrongful-looking denials, and patient harm. The piece argues the public still can't tell whether human reviewers or the data behind the AI are doing anything to prevent unfair denials.
- CMS launched WISeR six months after announcing it, and sent corrective action requests to at least four of six vendors over slow determinations, denials that looked inaccurate, and poor provider communication.
- One vendor denied more requests than it approved in the early months, and the payment structure gives vendors an incentive to deny claims, per the government's own words.
- "Human in the loop" safeguards may not work: reviewers can rubber-stamp AI output, and other insurers' systems have shown reviewers spending about 1.2 seconds per case.
EFF sued the government in March to get records on WISeR, a Medicare program that uses AI to review prior authorization requests for certain services. Its first release showed delays, denials, operational chaos, and reports of patient harm. This new batch, more than 800 pages, adds detail. It includes requests for information and corrective action plans CMS sent to four of the six vendors, the vendors' responses, applications from 66 companies, and contracts and task orders for the program. CMS launched WISeR six months after announcing it, despite providers asking for a delay, and CMS itself acknowledged an "extremely condensed timeline."
The letters show the rollout went badly across more than one vendor. CMS flagged "determination turnaround times that exceed established guidelines" and "denials that appear inaccurate or inconsistent with applicable criteria" at Virtix. It told Zyter, which runs on Palantir infrastructure, that it was not issuing determination letters on time. Innovaccer got two CAP requests and a "Performance Notice of Non-Compliance" for failing to deliver a working provider portal. Vendors pushed back by blaming the compressed timeline, and one said it joined a month after the others and hit "technical and workflow stabilization challenges." The notable part is that CMS apparently didn't disclose how widespread these problems were, even while staff were privately raising noncompliance with vendors. Fixes after the fact don't undo harms to patients and providers during those early months.
EFF's central argument is that the public still can't tell whether WISeR's safeguards work. CMS's answer to AI risk is that a human clinician must review any denial, and its FAQ says non-affirmations "cannot be performed solely by technology." EFF says that rule means little without details: how much time reviewers get, and whether they can actually overrule the software. Outside research backs the concern. Reviewers tend to rubber-stamp automated recommendations, one insurer reportedly pressured clinicians to accept its AI's calls, and another system let reviewers reject claims without opening the file, with reviewers averaging 1.2 seconds per case. Vendor payment structures, which CMS itself described as giving an "incentive to deny as many claims as possible," make this worse. One vendor touted human-in-the-loop workflows and audit trails, yet it denied more requests than it approved early on. EFF also wants clarity on what data feeds these decisions, since bad inputs can produce biased or inaccurate outcomes. EFF's ask is that Congress set firmer legal limits on automated decision-making in Medicare.