The AI That's Better Than Your Doctor Won't Save You Until Lawyers Get Involved

A study published yesterday in Science found that OpenAI's o1 model outperformed experienced emergency room physicians across five diagnostic tasks using real patient records from Beth Israel Deaconess Medical Center. "The model outperformed our very large physician baseline," Harvard's Raj Manrai told NPR¹. The headlines practically write themselves: AI beats doctors. But the headlines are pointing at the wrong problem.

The real inflection point in clinical AI isn't whether machines can outdiagnose humans. In specific, well-defined tasks, they demonstrably can. The real question, the one that will determine whether any of this actually reaches patients, is who gets sued when a hospital doesn't use a tool that would have caught the cancer, the sepsis, the embolism. And the answer, right now, is: nobody.

I think this is the most consequential gap in American health policy that almost no one is talking about. The current legal framework creates an asymmetry so perverse it's almost elegant in its dysfunction: hospitals face potential liability for adopting AI that makes an error, but face zero liability for declining to adopt AI that would have prevented a death. This is not a neutral equilibrium. It actively suppresses the deployment of potentially life-saving technology.

The evidence that AI outperforms doctors in certain diagnostic contexts is no longer speculative. The Harvard/Beth Israel study is just the latest data point. The MASAI randomized controlled trial², published in The Lancet in January 2026, enrolled over 105,000 Swedish women and found that AI-supported mammography achieved a 29% increase in cancer detection, 12% fewer interval cancers (cancers missed between screenings), and 6.7 percentage points higher sensitivity than standard double-reading by radiologists, all without increasing false positives. A separate nationwide implementation study in Germany³ (463,094 women, 12 sites) confirmed a statistically superior detection rate in real-world conditions. These aren't curated benchmarks. They're population-scale, prospective data.

And yet the legal system treats a hospital's decision not to deploy these tools as if it were ethically and legally inert.

The standard mechanism for forcing adoption of better medical tools has always been malpractice law. When a radiologist uses outdated equipment and misses a finding that current technology would have caught, that's already a recognized basis for negligence claims. The American Law Institute's newly approved Restatement (Third) of Torts: Medical Malpractice⁴ (voted through in May 2024, now being prepared for publication) formally shifts U.S. malpractice standards away from the old custom-based approach ("did the doctor do what other doctors do?") toward an evidence-based reasonableness standard that incorporates clinical guidelines and scientific evidence. As the Guirl Law Firm's analysis⁵ explains, the Restatement explicitly states that compliance with medical custom "is relevant but not decisive." A hospital's defense that "nobody else is using AI either" is losing its legal force, slowly but unmistakably.

The direction is clear. But I want to be precise about where we actually are, because the gap between where the law is headed and where it stands today is exactly where patients are dying.

The strongest case against rushing to impose non-adoption liability is that the underlying technology is unevenly validated. The FDA has now cleared 1,451 AI-enabled medical devices through end-2025⁶, nearly doubling since 2022. But according to a systematic review in JAMA Research Letter⁷, 97% of those were cleared via the 510(k) pathway, which requires only "substantial equivalence" to an existing device, not prospective clinical trials. Only about 5% of radiology AI devices underwent prospective testing. A 2025 JAMA Network Open study⁸ of 903 FDA-approved AI medical devices found clinical performance data was publicly available for only 55.9% at the time of clearance. That's an uncomfortable truth for anyone arguing hospitals should be penalized for not buying these products.

The equity dimension sharpens the problem further. A Nature Health study of 3,560 U.S. hospitals⁹ found that AI adoption is "considerably clustered, with hotspots and coldspots," and that "regions with greater healthcare access needs were less likely to have hospitals with AI-based predictive models." Rural hospitals, independent facilities, hospitals serving the most deprived neighborhoods: they adopt AI at systematically lower rates, not because they're negligent but because they lack the infrastructure, IT staff, and capital. Imposing blanket liability in this landscape doesn't close the gap. It punishes the institutions that can least afford it.

And there's the deskilling problem. ECRI designated "Navigating the AI Diagnostic Dilemma" as healthcare's number one patient safety concern for 2026¹⁰, warning that AI tools "can contribute to diagnostic mistakes" and "erode clinicians' critical thinking skills." This isn't theoretical anxiety. A controlled study found erroneous AI prompts increased false-positive mammography recalls by up to 12% among experienced radiologists, and a colonoscopy RCT found adenoma detection rates dropped significantly when physicians reverted to non-AI procedures after prolonged AI-assisted work.

So where does this leave us? Not, I think, with a "both sides" answer. The case for targeted, evidence-calibrated non-adoption liability is stronger than the case against it, but only if we're specific about what "targeted" means.

Here is the position I can defend: for well-resourced hospital systems, in diagnostic domains where prospective RCT or large-scale real-world evidence demonstrates AI superiority (mammography being the clearest current case), the failure to adopt validated AI tools should increasingly carry legal consequences. Not tomorrow, not through some novel tort theory that doesn't exist yet, but through the ordinary evolution of the standard of care as the ALI Restatement diffuses through state courts and as reimbursement frameworks catch up. The Milbank Quarterly's peer-reviewed analysis¹¹ already acknowledges that health systems face potential liability both for adopting AI they can't properly vet and for failing to adopt AI that improves care. That symmetrical exposure needs to be resolved, not deferred indefinitely.

At the same time, any liability framework that doesn't address three things is premature and dangerous: (1) safe harbors for resource-constrained hospitals that genuinely cannot implement AI given current infrastructure (malpractice doctrine already adjusts the standard of care to what is "reasonable in that environment"); (2) developer accountability, so that AI companies can't market aggressively while offloading all liability onto hospitals via the 510(k) pathway; and (3) governance requirements, so that adoption includes training, monitoring, and bias auditing, not just procurement.

The University of Michigan / NCBI legal analysis¹² captures the current dysfunction precisely: "malpractice liability concerns incentivise physicians to follow the standard of care they would have followed before, no matter what the AI suggests." That means the current legal framework doesn't just fail to encourage AI adoption; it actively discourages it. Every day this equilibrium persists is a day when the 250,000+ annual deaths from medical errors in the U.S. include some fraction that a validated AI tool would have prevented.

The indicator to watch is straightforward: the first U.S. malpractice lawsuit in which a plaintiff argues that a well-resourced hospital's failure to deploy AI-supported mammography (or an equivalent tool with RCT-level evidence) constituted a breach of the standard of care. That case is coming. The MASAI data, the ALI Restatement, and the sheer velocity of FDA clearances make it a matter of when, not whether. I expect it within 18 months. When it arrives, it will force every hospital procurement committee in the country to recalculate whether the cost of adopting AI is really higher than the cost of not adopting it.