The claim
Circulating September 2026
“Unitedhealth insurance used an AI with a 90% error rate to deny healthcare coverage; according to new lawsuit”
Most viral claims about litigation get the lawsuit wrong. This one quotes it correctly. The trouble starts in the same document, four dozen paragraphs later, where the complaint explains where its own number came from.
Verdicts
| Claim | Rating |
|---|---|
| A lawsuit alleges this, in these words | True |
| The 90% figure is an error rate | False |
| UnitedHealth used AI to deny coverage | Unproven |
| It is a “new lawsuit” | False |
The lawsuit does say this
The case is Estate of Gene B. Lokken et al. v. UnitedHealth Group, in the District of Minnesota. The operative pleading is the amended complaint filed in April 2024, and its first paragraph reads:
First Amended Complaint, Dkt. 34, ¶1
“This putative class action arises from Defendants’ illegal deployment of artificial intelligence (AI) in place of real medical professionals to wrongfully deny elderly patients care owed to them under Medicare Advantage Plans by overriding their treating physicians’ determinations as to medically necessary care based on an AI model that Defendants know has a 90% error rate.”[1]
So the phrase came from lawyers, not from a meme account. Across 59 pages the words “error rate” appear exactly twice, both in that opening passage.
And the lawsuit undercuts itself
The complaint offers one piece of support for the figure, at paragraph 47:
First Amended Complaint, Dkt. 34, ¶47
“Upon information and belief, over 90 percent of patient claim denials are reversed through either an internal appeal process or through federal Administrative Law Judge (ALJ) proceedings.”[1]
That is a reversal rate among denials that were appealed. It is not a rate across denials. The paragraph is pleaded “upon information and belief,” which is the formula lawyers use for something they believe but cannot yet prove, and it carries no citation at all. The footnote on that passage attaches to a separate figure about prior authorizations. No population is named, no time period, no insurer.
Six paragraphs later the complaint supplies the number that dismantles its own headline:
First Amended Complaint, Dkt. 34, ¶53
“Roughly 0.2% of consumers appeal their health insurance claim denials.”[1]
By the complaint’s own arithmetic, the 90% describes roughly one denial in five hundred, and those five hundred are not a random sample.
People who appeal are the ones with the strongest cases, the most persistence, and someone to help them. A high success rate among that group is what you would expect whether or not the underlying system is broken.
The 0.2% also comes from the wrong place. Its footnote cites a study of ACA marketplace plans in 2021.[7] This is a Medicare Advantage case about post-acute care. The complaint takes an appeal rate measured in one insurance market and applies it to another.
The problem underneath is real
None of that means the concern is invented. A federal auditor has since measured the same phenomenon, named the same contractor, and found a number higher than the one in the lawsuit.
HHS Office of Inspector General, OEI-09-24-00331, June 2026
“The MAO contractor naviHealth processed half of all requests for SNF admission and denied 14 percent of them—a higher denial rate than MAOs that processed requests internally (11 percent) and other contractors (9 percent).”
“MAOs later overturned 97 percent of SNF denials issued by naviHealth when enrollees appealed.”[5]
Ninety-seven percent is higher than ninety. It is also the same kind of number, and the auditor says so in the same breath: enrollees appealed 18 percent of skilled nursing denials, and the 97 percent describes only those. The other 82 percent of denials were never tested by anyone. The report’s own title says these figures raise concerns about initial denials, which is an inference rather than a finding.
So the 97 percent is not a correction to the 90 percent. It is the same measurement, taken properly, on a narrower population, by someone with subpoena power. Neither number tells you how often the tool was wrong.
For the measured version, the Inspector General sampled denied Medicare Advantage requests and checked them against Medicare coverage rules:
HHS Office of Inspector General, OEI-09-18-00260, 2022
“We found that among the prior authorization requests that MAOs denied, 13 percent met Medicare coverage rules—in other words, these services likely would have been approved for these beneficiaries under original Medicare.”[6]
A Senate subcommittee also documented a sharp rise. Its 2024 staff report found UnitedHealthcare’s initial denial rate for post-acute prior authorizations went from 8.7 percent in 2019 to 22.7 percent in 2022, and for skilled nursing admissions from 1.4 percent to 12.6 percent.[8] Two cautions belong with those figures. It is a majority staff report rather than a bipartisan committee finding, written when the subcommittee had a different chairman. And it expressly declined to attribute the increases to software, saying the data is troubling “regardless of whether the decisions reflected in the data were the result of predictive technology or human discretion.”
What the court has and has not found
In February 2025 the judge ruled on the motion to dismiss. Five of the seven claims were dismissed with prejudice as preempted by the Medicare Act: unjust enrichment, insurance bad faith, and three state statutory claims. Two survived.[2]
What survived is narrower than the headlines suggest, and it is not about accuracy at all. The coverage documents told members this:
Evidence of Coverage, quoted at Dkt. 91
“UnitedHealthcare’s Clinical Services Staff and Physicians make decisions on the health care services you receive based on the appropriateness of care and service and existence of coverage.”[2]
The two live claims are breach of contract and breach of the implied covenant, and they survived precisely because deciding them requires only reading that sentence against what the company did. The question is whether UnitedHealth promised humans and delivered software, not whether the software was accurate.
At this stage a court assumes the allegations are true. It has made no finding about the 90 percent figure, and the phrase “error rate” does not appear anywhere in the 24-page ruling. UnitedHealth denies the allegations formally, in an answer filed this month,[4] and told reporters in 2023 that the tool “is not used to make coverage determinations.”[12]
The case is also not new. It was filed in November 2023, nearly three years ago. No class has been certified, the motion to certify one is not due until the end of this year, and the trial-ready date is December 2027.[3]
The rule that was proposed and withdrawn
The part of this story with the least coverage is what the government did about it.
In February 2024 Medicare regulators issued guidance saying an algorithm may help predict a length of stay, but that prediction “alone cannot be used as the basis to terminate post-acute care services.”[10] That is guidance interpreting a rule about individualized decisions. The regulations themselves never mention algorithms or artificial intelligence.
In December 2024 the agency proposed to fix that, with a provision titled “Guardrails for Artificial Intelligence” that would have written the requirement into the rules.[9] In April 2025 it dropped the proposal:
Medicare Advantage final rule, 90 Fed. Reg. 15792 (15 April 2025)
“We also do not intend to finalize the following provisions from the proposed rule: … and Ensuring Equitable Access to Medicare Advantage Services-- Guardrails for Artificial Intelligence (AI).”[11]
The agency added that it would “continue to consider” rulemaking in the area. That sentence is the only time artificial intelligence is mentioned in the entire final rule, and no binding provision governing algorithmic coverage decisions has been adopted since.
Which leaves the accurate version of the story, and it is not smaller than the viral one. A federal auditor found that nearly every appealed denial from this contractor was overturned. Almost nobody appeals. A judge is letting a case proceed on the theory that an insurer promised human decisions and allegedly delivered software. And the rule that would have required humans by name was proposed, then withdrawn.
Who says what
Five parties have taken a position on whether an algorithm denied care. None of them agree, and only two have measured anything.
| Party | Position | Basis |
|---|---|---|
| Plaintiffs | AI replaced physicians and has a 90% error rate | Allegation, pleaded on information and belief |
| UnitedHealth | The tool does not make coverage determinations; denies the allegations | Formal answer and reply brief |
| HHS Inspector General | 97% of appealed naviHealth denials overturned; 13% of audited denials should have been approved | Two audits, sampled data |
| Senate subcommittee | Denials rose sharply; declines to say whether software caused it | Majority staff report, insurer data |
| The court | No finding either way; two contract claims may proceed | Allegations assumed true at this stage |
What to watch
The case is live and three dates decide what it becomes.
| Date | What happens |
|---|---|
| 30 September 2026 | Parties must report to the court on settlement discussions. A settlement would end the case without any finding about the algorithm. |
| 31 December 2026 | Motion for class certification due. No class has been certified yet, and denial would shrink this to a handful of individual claims. |
| December 2027 | Trial-ready date. The earliest any fact about nH Predict could be established rather than alleged. |
Discovery closes in March 2027, and the documents ordered produced in March 2026 are the first time anyone outside the company will see how the tool was built and used. The source code stays out of it.
Sources
- First Amended Class Action Complaint, Dkt. 34
- Memorandum Opinion and Order on Motion to Dismiss, Dkt. 91
- Order on Motion to Compel Discovery, Dkt. 162
- Second Amended Answer to Amended Complaint, Dkt. 206
- Medicare Advantage Organizations Overturned Nearly All Appealed Prior Authorization Denials for Skilled Nursing Facility Admission
- Some Medicare Advantage Organization Denials of Prior Authorization Requests Raise Concerns About Beneficiary Access to Medically Necessary Care
- Claims Denials and Appeals in ACA Marketplace Plans in 2021
- Refusal of Recovery: How Medicare Advantage Insurers Have Denied Patients Access to Post-Acute Care
- Medicare Advantage proposed rule, 89 Fed. Reg. 99340
- Frequently Asked Questions related to Coverage Criteria and Utilization Management Requirements
- Medicare Advantage final rule, 90 Fed. Reg. 15792
- AI with 90% error rate forces elderly out of rehab, nursing homes, suit claims
