Your AI Will See You Now
AI is 4x more accurate than human doctors, but patients aren't reaping the benefits
The next time you're sitting on that paper-covered exam table, your doctor won’t tell you the most important fact about your visit: They’re weighing their decades of training against an AI that’s four times more accurate—80% (AI) to their 20% (human) on complex cases. And they can’t tell you whether they trusted their own experience or a machine’s with your life.
That four-fold difference isn’t a projection or possibility. It’s the result of Microsoft's study on 56 of medicine's hardest cases—the diagnostic puzzles regularly published by the New England Journal of Medicine that humble even veteran physicians. When tested head-to-head, the AI didn’t just edge out human performance. It exposed a truth that could change everything about health care and the role of medical providers within it.
In fact, it’s already happening. According to a new American Medical Association survey, 66% of physicians now use AI—up 78% from last year. Two out of three doctors are making decisions with technology that dramatically outperforms human performance. Yet our entire legal system still assumes that the “reasonable physician” standard means a human physician. Should it?
The Paradox
Imagine a patient presenting in a doctor’s office with unusual symptoms that the doctor can’t quite place. Plugging the information into an AI system, the machine predicts the patient has a rare autoimmune condition with a confidence level of 80%. The doctor, trained to recognize common diagnoses, suspects something simpler, just a common cold triggering a post-viral sequelae. Which prediction governs the patient’s treatment? And who’s liable if either is wrong?
In their April 2024 guidance, the Federation of State Medical Boards (FSMB), the organization that provides legislative advocates for state medical boards, tried to answer this question. Their solution proposed that physicians remain fully responsible for all AI-assisted decisions and exercise their independent clinical judgment when using AI tools.
Sit with that idea. Does it really make sense for physicians to assume complete liability for technology that outperforms them by 400%? And then defend the decisions made by systems that are so complex that physicians can’t even explain how they arrived at their recommendations, nor why their own clinical judgment is inferior?
An MIT Media Lab study this month adds another layer to this conundrum. Regular AI use creates temporary but measurable changes in brain activity. Participants using ChatGPT showed “weaker neural connectivity and under-engagement of alpha and beta networks.” If AI use detrimentally affects cognitive patterns, which seems to be verifiably the case, how can physicians maintain the “independent judgment” the FSMB requires?
At Aspen Ideas Health last week, where I served as a panelist on the Artificial Intelligence Revolution, this collision between capability and liability dominated the conversation. When I asked the packed room of healthcare leaders how many had used ChatGPT for medical questions, hands shot up everywhere. When I then asked about uploading medical records, the nervous laughter said everything.
My fellow panelist Micky Tripathi from Mayo Clinic underscored the gridlock. Mayo has AI tools ready to deploy today that could help patients immediately. But they can’t use them. Their governance processes can’t evaluate technology that evolves monthly while liability questions remain unanswered.
Innovation vs. Regulation
During our panel, Karen DeSalvo from Google captured the fundamental mismatch. Innovation races at Silicon Valley speed. Regulation crawls at government pace. By the time lawmakers and courts answer today’s questions, the technology will be three generations ahead.
In the Q&A that followed, a physician asked about using AI to solve healthcare’s most maddening problem—fragmented medical records across different systems. The technology exists today to unify your entire medical history instantly. But the regulatory framework? The liability questions? Unanswered. So fragmentation continues, information stays siloed, and patients suffer while experts, policymakers, and lawyers debate.
These unanswered questions multiply daily. Will malpractice plaintiffs argue that failing to use an 80% accurate tool equals negligence? Or will they claim that following AI recommendations—even demonstrably superior ones—breaches the standard of care by abandoning independent judgment?
The FSMB suggests falling back on “established ethical principles,” but how can we fall back on outdated ethical frameworks (designed for human-only medicine) to solve a fundamentally new problem where the numbers show humans are vastly outperformed? It’s like trying to use horse-and-buggy traffic laws for highways with cars. The old framework simply doesn’t fit the new reality.
Some propose making AI companies liable, but these are just tools—sophisticated calculators. Others suggest new insurance frameworks, but you can’t price risk that hasn’t been defined. The boldest proposals would abandon fault-based malpractice entirely for AI-assisted care, creating no-fault compensation funded by AI’s cost savings, where some combination of healthcare providers and hospitals would pay into a fund, AI companies might contribute based on their tools’ usages, insurance companies could be required to redirect some malpractice premiums to the fund, or government subsidies could contribute from overall health savings.
But none of these proposals address the fact that we’re asking physicians to maintain skills that machines have already surpassed while simultaneously warning them against the “dependence and skill degradation” that comes from using these very machines and the superior diagnostic capabilities they provide.
Your next doctor's visit will happen in this gap between mathematical reality and legal theory. Your physician will choose between human judgment and machine analysis with undefined consequences. They’ll make decisions that could save your life or end their career—or both.
Courts haven’t yet ruled whether there’s a duty to use AI when it’s superior to human judgment. They haven't decided if following AI recommendations while maintaining independent judgment is even possible. These are daily realities facing physicians, where doing the medically right thing and the legally safe thing are starting to conflict.
Microsoft didn’t intend to expose this crisis. They simply built something that diagnoses complex cases far better than physicians do. But that achievement revealed an uncomfortable truth: Medical malpractice law assumes human judgment defines reasonable care. When machines obliterate that assumption, the question isn’t whether the framework will change. It’s how many patients will be affected before it does.
The head of Microsoft’s AI tech unit, Mustafa Suleyman, predicts that these systems will be “almost error-free in the next 5-10 years.” Will the law catch up to the math?
The AI will see you now. Whether the law will protect anyone in the exam room remains dangerously unclear.
 | A guest post by
|
An important and wise discussion, BUT written by a good lawyer, making many good observations and omitting others!
I am speaking as a family doc with 40 years of clinical experience, 10 years of medical bioethics experience, and 8 years on a state medical licensing board (much more pertinent than the FSMB's dictums). Your discussion has 2 practical and logical omissions: (1) When you say "failing to use an 80% accurate tool [may] equals negligence," the '80% better' applies only to a small (unmentioned, undetermined and unspecified) number of medical encounters for RARE, complex conditions. Most encounters are 'ordinary' and experienced docs and AI would likely perform equally well. The real question is how to decide when docs and patients will both benefit from using the (occasionally hallucinating) black box of AI for a better outcome. Using AI for every encounter is not a good prospect because that will certainly lead to an atrophy of human clinical skills. (2) In any case the goal of medical care is precise diagnosis and best treatment with optimal patient outcome. Whether the path is taken by personal clinical wisdom and experience and/or AI, the choice to follow the differential diagnosis and what interventions to make should still rest in the mind of a human being. And in medicine, the results (and feedback) are clear, explicit, and proximate - the patient improves or does not. The consequences are not only not theoretical, they can be grave.
Clinical medicine learned to deal with radiographs, blood tests, genomics; technological tools will continue to evolve, but compassionate, wise, humane judgment will remain central in medicine's essence for the care of people.
This study doesn’t quite provide what’s needed, though. Maybe this has already been done, but there should be a double-blind classic study — if that’s still possible, since AI is so widely used— with doctors not using AI and doctors relying on AI, for a broad range of ordinary, run of the mill cases.
The classic medical advice to young medical students and doctors is not to diagnose “zebras” (rare diagnoses) unless they have to, as IRL most cases are horses, not zebras. The study Microsoft did is just the opposite, designed exclusively to look at zebras, and rare ones at that — not what really matters for improving health care.