Lawsuits over AI coverage denials test what a human reviewer's sign-off is worth

Insurers have scored policyholders with algorithms for decades ^[6]. The older ones took inputs such as age, weight and smoking status and returned a rough estimate of the care a person would need before death ^[6]. Vox calls that approach simple and comparatively transparent ^[6]. On a claim or a prior authorization request, a human doctor was supposed to review the file for the insurer and make the final decision ^[7]. According to Daniel Schwarcz, a University of Minnesota law professor who has studied insurers' AI use, the newer programs make very specific predictions about a person's future health, and those predictions are used to approve or deny claims ^[8]. What goes into them is often opaque from outside. So is whether the model or the person holds final authority ^[8]. Schwarcz described the difficulty directly. "The reason we want human review is we want human substantive judgment. We want AI to be supporting that," he said. "But the difficulty is distinguishing between AI supporting meaningful human processing and AI basically replacing the human with the human just passively defaulting to the AI judgment. It's just incredibly hard to distinguish between those two in practice, particularly in rules or regulations." ^[9] The board-deck version of this risk is short: a clinician signs every adverse decision, so the company is covered. That version is incomplete because a signature shows only that a person was present. UnitedHealth is accused of denying payment for care that patients' doctors said was necessary ^[1]. The complaints describe a patient who may need months of rehab being cut off by an algorithm after a couple of weeks ^[5]. A reviewer's sign-off that matches the model in those cases is the situation Schwarcz says is hard to tell apart from passive defaulting ^[9]. The incentive runs one way. Payers, Vox notes, have a business incentive to pay as little as they can ^[11]. A skeptic would say these are allegations, and no insurer has lost anything yet. On the record, that is correct. Humana and Cigna face their own litigation over similar claims ^[2]. The suits remain unsettled, and policymakers have not worked out how to regulate insurers' use of the technology ^[12]. The reporting includes no ruling, damages figure or settlement. The evidence establishes legal exposure at three insurers ^[14]. Whether that exposure becomes liability for overriding a doctor with a model depends on rulings still to come ^[12]. The federal government buys this technology as well as potentially regulating it. Medicare has launched its own AI tool to help decide whether to approve prior authorizations ^[3]. Doctors, meanwhile, are sending the federal government reports of patients in tears because they cannot get medicine to manage chronic pain ^[4]. With the government adopting the practice it would have to police, I'd expect the working definition of adequate human review to come out of these lawsuits first. The complaints turn on the gap between what a model takes in and what a patient's own doctor saw. "There really are a lot of factors that really can't be quantified and so aren't really suitable to go into these models. If we start to rely exclusively on the models, I think we run the risk of losing that human touch," said Mika Hamer, a health services researcher at the University of Maryland ^[10]. Vox tells patients to appeal, reporting that most claims are overturned on appeal ^[13]. Where that holds, a company's own appeal records would show reviewers reversing most of the first answers that patients contest. This quarter's choice is what a review record contains. A record of what the reviewer examined, and where they departed from the model, gives a company something to set against the treating doctor's opinion. A log of sign-offs and timestamps cannot separate meaningful review from the passive defaulting Schwarcz describes ^[9].