AI-Governance Claims, Checked Against Primary Sources
Why this report exists
Most claims about AI never get independently checked. They get repeated. This report turns the series toward a live, high-stakes deployment: the automated and algorithmic tools used to deny health coverage in Medicare Advantage.
Third in a standing series. We take claims circulating widely in the market and check each against the primary source. Here, that source is the U.S. Department of Health & Human Services, Office of Inspector General (HHS-OIG) — official government audits, not litigation, advocacy, or press. Where a figure reflects a specific audit period, we say so; an audit of 2019 denials is not a claim about today. Two disciplines govern the report:
- We verify claims, not claimants. No company is named. We are not grading anyone's product or character — only whether a statement holds against the record.
- Precision over posture. The goal is not to "debunk." It is to state the exact position, including the nuance. An overcorrection is just a new error.
Verdict scale · VERIFIED / PARTIALLY VERIFIED / NOT VERIFIED
HHS-OIG audited a stratified random sample of denials from the 15 largest Medicare Advantage organizations (OEI-09-18-00260, issued April 27, 2022; denials drawn from June 2019). Among prior-authorization requests the plans denied, 13% actually met Medicare coverage rules — the care likely would have been approved under Original Medicare. Among payment requests denied, an estimated 18% met both Medicare coverage rules and the plan's own billing rules.
The rule — MA plans must cover at least what Original Medicare covers — is real. But compliance is not automatic: a measurable share of denials in the sample failed Medicare's own standard. "Every denial reflects Medicare's rules" is not supported by the record. (Scope: an audit of 2019 denials published in 2022, not a current-year rate.)
HHS-OIG (OEI-09-16-00410, September 2018; 2014–2016 data) found that when denials were appealed, Medicare Advantage organizations overturned 75% of their own denials — roughly 216,000 overturned denials per year. But in the same period, only 1% of denials were ever appealed. Separately, CMS audits cited 56% of audited MA contracts for inappropriate denials and 45% for sending incomplete or incorrect denial letters.
The appeals mechanism does work — it reverses most denials that reach it. That is exactly why the 1% appeal rate matters: a correction process that overturns three of every four challenges, but is triggered on one in a hundred denials, is not a safety net the system leans on. High overturn plus low appeal is not reassurance — it is the shape of a problem.
A June 2026 HHS-OIG post-acute audit (OEI-09-24-00331, published June 11, 2026; June 2024 data across 19 MA organizations) found that when skilled-nursing-facility prior-authorization denials were appealed, plans overturned 95% of them in favor of the enrollee. For the single review contractor that processed roughly half of all SNF admission requests — the largest user of a predictive post-acute tool in the market — 97% of its appealed denials were overturned, against a 14% initial denial rate. Only 18% of SNF denials were appealed at all.
"Faster" may hold; "more accurate" does not on this evidence. A decision process whose denials are overturned 95–97% of the time on appeal is not demonstrating accuracy — it is demonstrating that the initial "no" rarely survives independent review. The claim that predictive tooling improves accuracy in this setting is unsupported by the audited outcomes. (We name no vendor or plan; the contractor and rates above are as characterized in the OIG report.)
The pattern — the Verification Gap
Three claims that shape how a high-stakes AI deployment is defended, checked against the government's own audits: one true-but-incomplete, one true-but-rarely-triggered, one unsupported as stated.
The thread is the gap between the reported account of these systems — "denials reflect the rules," "appeals catch the errors," "the tools improve accuracy" — and the independently verified record: a measurable share of denials failed Medicare's own rules; the appeal that would catch them is filed 1% of the time; and where it is filed, 95–97% of post-acute denials are reversed. That gap — reported versus independently verified — is the whole reason independent verification exists.
None of this is an accusation against any company. It is the public audit record, read precisely. Self-report and repetition are not verification.
What this report is — and isn't
- It is an independent check of public claims against primary sources, issued at no charge.
- It is not legal advice, and not a compliance opinion for any organization. Confirm your own obligations with counsel and the current record.
- It names no company and evaluates no product. It grades statements, not people.
MarginSignal OS — independent verification of what deployed AI, and the claims around it, actually do.
Corrections to any finding in this report are welcome and will be issued publicly. That is the point.