California a model for needed regulation of AI in prior authorization administration

Recent events regarding the assassination of the CEO of UnitedHealthcare have brought to the forefront widespread patient frustrations with the state of the American health insurance system.

As a scholar of health insurance barriers, the last week has admittedly been rather strange for me.

Murder is not justified, and vigilante justice is not justice. And while some have felt comfortable expressing feelings of vindication in their exasperation if not fury toward private insurance, this act is almost assuredly not going to lead a shift in policy by UHC and other private insurers. These are widespread problems that require systemic interventions.

There is a reason why only 31% of Americans trust the American health care system, and my forthcoming book Coverage Denied: How Health Insurers Drive Inequality in the United States (Cambridge University Press, 2026) sheds light on many aspects of this problem.

Health care in the United States is very expensive for patients to obtain. We spend 1 out of every 6 dollars in the U.S. on health care, without a great bang for buck. What’s more, navigating health insurance has a great deal of administrative burden for both patients and their physicians (the joys of prior authorization!), even when things are ultimately covered under one’s health insurance policy. And then there are the coverage denials, which I show wreak havoc on patients’ health and economic lives, with the worst burdens shouldered by those from marginalized groups.

One of the recent realms of consternation with private health insurance is the deployment of AI tools for prior authorization processing. On the one hand, there are obvious efficiency gains for health insurers (and theoretically, for patients, if there aren’t errors in processing… a major caveat). After all, insurers can inexpensively and quickly process large volumes of prior authorization requests, helping patients to expeditiously access their prescribed care if approved.

You can probably guess the problems, however.

Health insurers including UnitedHealthcare, Cigna, and others use for prior authorization processing AI tools (PxDx and nH Predict) that according to class action litigation yield a reversal rate of a whopping 90% upon appeal.

So, how do these tools work? In a nutshell, they analyze millions of datapoints related to the patient’s condition and prescribed course of care, to generate recommendations based on similar patients.

In the 2024 report Refusal of Recovery: How Medicare Advantage Insurers Have Denied Patients Access to Post-Acute Care, led by Sen. Richard Blumenthal (D-CT), the Permanent Subcommittee on Investigations observed that between 2019 and 2022, UnitedHealthcare, Humana, and CVS denied prior authorization requests for post-acute care at far higher rates than they did other types of care, thus keeping post-acute care out of reach for far too many Medicare Advantage enrollees. Indeed, UnitedHealthcare’s overall prior authorization denial rate increased from 16.3% to 22.7% after approving the use of AI in this processing – but the denial rate for skilled nursing facility care was nine times higher in 2022 than in 2019.

In The Estate of Gene B. Lokken and The Estate of Dale Henry Tetzloff, individually and on behalf of all others similarly situated v. UnitedHealth Group, Inc., UnitedHealthcare, Inc., and NaviHealth, Inc., the plaintiffs that UnitedHealth’s deployment of this AI tool led them to “wrongfully deny elderly patients care owed to them under Medicare Advantage plans by overriding their treating physicians’ determinations as medically necessary care” despite the known “90% error rate" of the “flawed AI model” of nh Predict. The plaintiffs alleged that this high level of wrongful denials stems from the model’s “rigid and unrealistic predictions for recovery,” such that while Medicare Advantage enrollees are entitled to 100 days of skilled nursing facility care following a three-day (or longer) hospitalization, “with the use of the nH Predict AI Model, Defendants cut off payment in a fraction of that time. Patients rarely stay in a nursing home more than 14 days before they start receiving payment denials.” Frustration surrounding this is only exacerbated by the lack of transparency surrounding these determinations, given UnitedHealth’s assertion that the decisionmaking rationale is proprietary.

Humana and Cigna are facing similar class action lawsuits concerning their use of nH Predict and PxDx respectively in the bulk processing and denial of prior authorizations and claims, with Cigna’s reliance on AI earning attention from investigative reporting by ProPublica.

Does this actually mean it’s a 90% error rate, as the class action lawsuits allege? Not necessarily. It’s not hard to imagine that there is some unknown number of “go away” reversals – that is, conditions in which it is more costly to continue to challenge the coverage determination than it is to pay for the prescribed care, “medically necessary” or not (and there is discretion and politics involved in medical necessity determinations, as political scientist Dan Skinner highlights in Medical Necessity: Health Care Access and the Politics of Decision Making).

But there’s another important wrinkle: Many people do not appeal coverage denials for a variety of reasons including informational barriers and broader administrative burdens given high health literacy demands, and I find in my own research that people from marginalized groups are – not surprisingly – less likely than their more affluent and White counterparts to do so and to do so successfully. Thus, while many patients are successful upon appeal, the increase in wrongful denials not only induces some delays in care, but it exacerbates the already prevalent administrative burdens in the American health care system (with particular attention to Medicare Advantage enrollees, a population that given more advanced age would be expected to have somewhat lower health literacy). Such denials in turn disrupt the health and economic lives of these patients and their families.

The California state legislature, led by state Sen. Josh Becker, has taken action on this subject, passing unanimously in the state senate and nearly unanimously in the state assembly S.B. 1120 (the Physicians Make Decisions Act) to require that effective January 1, 2025, which Gov. Gavin Newsom subsequently signed into law. According to the Physicians Make Decisions Act, only a licensed physician or health care professional competent to make clinical decisions in that particular area of health care may make a medical necessity determination; no AI program or algorithm may deny, delay, or modify health care services based on medical necessity, and utilization management must be on the insured’s own medical record rather than solely on group data sets; and AI systems used by health plans are subject to audit and compliance reviews by the California Department of Managed Health Care.

In its analysis of the bill, the California Assembly Committee on Privacy and Consumer Protection drew attention to a deeply troubling example of AI-induced error in claim processing: “a hospital trained AI models on a dataset of 15,000 pneumonia patients in order to develop a model that could identify which pneumonia patients were at the greatest risk in order to triage new patients. During testing, it was discovered that one of the most accurate models recommended outpatient status for asthmatics – a life-threateningly dangerous error based on a statistical correlation: asthmatics are less likely to die from pneumonia than the general population precisely because asthma is such a serious risk factor that asthmatics automatically get elevated care.” Wrongful determinations aren’t always mere inconveniences. They can have life or death consequences.

Speaking on the legislation, Sen. Becker said, “Artificial intelligence is an important tool in diagnosing and treating patients, but it should not be the only factor in determining if or what kinds of health care a patient receives,” given that an algorithm will not be able to fully appreciate the complexity of an individual patient case. The legislation received the backing of the California Medical Association, the California Hospital Association, UC Davis Health, and a wide array of physician associations such as the California Academy of Family Physicians, California Chapter of American Cardiology, Physician Association of California, and Psychiatric Physicians Alliance of California, while earning opposition from the California Association of Health Plans and the California Association of Life and Health Insurance Plans.

Of course, this legislation is not costless. The California Department of Managed Health Care estimated that the legislation could result in higher premiums. What’s more, while the implementation of such protections could result in fewer wrongful denials being issued – which drives up patient and physician administrative burden – reinstating more reliance on physician determinations rather than technological tools can slow down processing for the patient seeking approval of their medication, scan, etc.

But whether you’re the senior in need of care in a skilled nursing facility or the asthmatic in need of hospital care for the treatment of pneumonia, a modest delay to yield a correct result is a reasonable price to pay.

Some members of Congress are taking note. Rep. Ami Bera (D-CA) and the House Task Force on AI (on which he sits) have reviewed the California legislation and will be watching to see the impact of the law’s implementation starting in January, noting that if it has the intended effects, “it could be a model nationwide.”

This is all in the setting of renewed attention to the challenges posed by prior authorization, with Sen. Roger Marshall (R-KS) reintroducing the Ensuring Seniors’ Timely Access to Care Act, which passed the House of Representatives with a voice vote in September 2022. And with the likely acceleration of reliance on Medicare Advantage (in which 99% of enrollees have prior authorization requirements for at least some care, a sharp contrast with traditional Medicare) under the incoming administration, tackling this issue will be all the more important.

Of course, we do not really know what the future holds for health reform in this space in light of the results of the 2024 election. While prior authorization reform has garnered bipartisan support both in California and nationally, it may not be a legislative priority. We will know more in the coming month whether the ongoing public outcries over these prevalent health insurance barriers will resonate with the incoming Congress.