How artificial intelligence controls your health insurance coverage

Over the past decade, health insurance companies have increasingly embraced the use of artificial intelligence algorithms. Unlike doctors and hospitals, which use AI to help diagnose and treat patients, health insurers use these algorithms to decide whether to pay for health care treatments and services that are recommended by a given patient’s physicians.

One of the most common examples is prior authorization, which is when your doctor needs to receive payment approval from your insurance company before providing you care. Many insurers use an algorithm to decide whether the requested care is “medically necessary” and should be covered.

These AI systems also help insurers decide how much care a patient is entitled to — for example, how many days of hospital care a patient can receive after surgery.

If an insurer declines to pay for a treatment your doctor recommends, you usually have three options. You can try to appeal the decision, but that process can take a lot of time, money and expert help. Only 1 in 500 claim denials are appealed. You can agree to a different treatment that your insurer will cover. Or you can pay for the recommended treatment yourself, which is often not realistic because of high health care costs.

As a legal scholar who studies health law and policy, I’m concerned about how insurance algorithms affect people’s health. Like with AI algorithms used by doctors and hospitals, these tools can potentially improve care and reduce costs. Insurers say that AI helps them make quick, safe decisions about what care is necessary and avoids wasteful or harmful treatments.

But there’s strong evidence that the opposite can be true. These systems are sometimes used to delay or deny care that should be covered, all in the name of saving money.

A pattern of withholding care

Presumably, companies feed a patient’s health care records and other relevant information into health care coverage algorithms and compare that information with current medical standards of care to decide whether to cover the patient’s claim. However, insurers have refused to disclose how these algorithms work in making such decisions, so it is impossible to say exactly how they operate in practice.

Using AI to review coverage saves insurers time and resources, especially because it means fewer medical professionals are needed to review each case. But the financial benefit to insurers doesn’t stop there. If an AI system quickly denies a valid claim, and the patient appeals, that appeal process can take years. If the patient is seriously ill and expected to die soon, the insurance company might save money simply by dragging out the process in the hope that the patient dies before the case is resolved.

This creates the disturbing possibility that insurers might use algorithms to withhold care for expensive, long-term or terminal health problems, such as chronic or other debilitating disabilities. One reporter put it bluntly: “Many older adults who spent their lives paying into Medicare now face amputation or cancer and are forced to either pay for care themselves or go without.”

Research supports this concern – patients with chronic illnesses are more likely to be denied coverage and suffer as a result. In addition, Black and Hispanic people and those of other nonwhite ethnicities, as well as people who identify as lesbian, gay, bisexual or transgender, are more likely to experience claims denials. Some evidence also suggests that prior authorization may increase rather than decrease health care system costs.

Insurers argue that patients can always pay for any treatment themselves, so they’re not really being denied care. But this argument ignores reality. These decisions have serious health consequences, especially when people can’t afford the care they need.

Moving toward regulation

Unlike medical algorithms, insurance AI tools are largely unregulated. They don’t have to go through Food and Drug Administration review, and insurance companies often say their algorithms are trade secrets.

That means there’s no public information about how these tools make decisions, and there’s no outside testing to see whether they’re safe, fair or effective. No peer-reviewed studies exist to show how well they actually work in the real world.

There does seem to be some momentum for change. The Centers for Medicare & Medicaid Services, or CMS, which is the federal agency in charge of Medicare and Medicaid, recently announced that insurers in Medicare Advantage plans must base decisions on the needs of individual patients – not just on generic criteria. But these rules still let insurers create their own decision-making standards, and they still don’t require any outside testing to prove their systems work before using them. Plus, federal rules can only regulate federal public health programs like Medicare. They do not apply to private insurers who do not provide federal health program coverage.

Some states, including Colorado, Georgia, Florida, Maine and Texas, have proposed laws to rein in insurance AI. A few have passed new laws, including a 2024 California statute that requires a licensed physician to supervise the use of insurance coverage algorithms.

But most state laws suffer from the same weaknesses as the new CMS rule. They leave too much control in the hands of insurers to decide how to define “medical necessity” and in what contexts to use algorithms for coverage decisions. They also don’t require those algorithms to be reviewed by neutral experts before use. And even strong state laws wouldn’t be enough, because states generally can’t regulate Medicare or insurers that operate outside their borders.

A role for the FDA

In the view of many health law experts, the gap between insurers’ actions and patient needs has become so wide that regulating health care coverage algorithms is now imperative. As I argue in an essay to be published in the Indiana Law Journal, the FDA is well positioned to do so.

The FDA is staffed with medical experts who have the capability to evaluate insurance algorithms before they are used to make coverage decisions. The agency already reviews many medical AI tools for safety and effectiveness. FDA oversight would also provide a uniform, national regulatory scheme instead of a patchwork of rules across the country.

Some people argue that the FDA’s power here is limited. For the purposes of FDA regulation, a medical device is defined as an instrument “intended for use in the diagnosis of disease or other conditions, or in the cure, mitigation, treatment, or prevention of disease.” Because health insurance algorithms are not used to diagnose, treat or prevent disease, Congress may need to amend the definition of a medical device before the FDA can regulate those algorithms.

If the FDA’s current authority isn’t enough to cover insurance algorithms, Congress could change the law to give it that power. Meanwhile, CMS and state governments could require independent testing of these algorithms for safety, accuracy and fairness. That might also push insurers to support a single national standard – like FDA regulation – instead of facing a patchwork of rules across the country.

The move toward regulating how health insurers use AI in determining coverage has clearly begun, but it is still awaiting a robust push. Patients’ lives are literally on the line.

Jennifer D. Oliva currently receives funding from NIDA to research the impact of pharmaceutical industry messaging on the opioid crisis among U.S. Military Veterans. She is affiliated with the UCSF/University of California College of the Law, San Francisco Consortium on Law, Science & Health Policy and Georgetown University Law Center O’Neill Institute for National & Global Health Law.