When physicians and AI work together, who is accountable? How to lay out medical liability

July 29, 2026.

The question of accountability in this new partnership between physician and machine presents itself to me as a problem of classification - not unlike the sorting of specimens whose characteristics seem to blend across categories. One reads of a proposed staging system, based on the magnitude of the machine’s role in patient care, and I confess my mind turns at once to the difficulty of drawing clear lines where influence is gradual and interwoven. The physician consults the tool; the tool suggests; the physician approves or modifies; the outcome follows. Where, in that sequence, does responsibility come to rest? It puts me in mind of the way we once classified species - as if nature respected our categories. We now know better. The boundaries are rarely sharp. I observe that we are attempting to impose a discrete staging upon what is likely a continuum of collaboration, and I cannot help but feel that the cases which will prove most troublesome are precisely those that fall between the stages - where the machine’s suggestion was persuasive but not determinative, where the physician’s judgment was assisted but not replaced. The anomalies, as always, will be the most instructive. It appears we are building a new kind of fossil record - one of decisions and errors - and the strata of liability will take decades to settle into something like a clear pattern. For now, I suspect we must proceed with the humility that comes from recognizing that we are still collecting the evidence.

The question of accountability when physicians and AI collaborate is rather like asking who is to blame when a carriage overturns - the driver, the horse, or the road? The answer, of course, depends on whether the reins were slack, the beast was spooked, or the stones were loose. But in medicine, we have long known that responsibility is not so easily parsed.

I recall a case from my own practice - a young woman with a fever that defied diagnosis. The senior physician declared it typhoid with the confidence of a man who had never been wrong. The junior, who had spent nights at her bedside, suspected something else. Had there been an “AI” in the room, would its verdict have been trusted over the junior’s hesitation? Or would the senior have seized upon its output as vindication?

The proposal of a “staging system” for AI’s role in diagnosis strikes me as sensible - if we must have these mechanical oracles, let us at least measure their influence. But I fear the real danger lies not in the machine’s error, but in the physician’s pride. A tool is only as good as the hand that wields it, and if the hand is too proud to question itself, the fault is not in the steel.

The profession has always resisted the notion that its judgment could be fallible. Puerperal fever was once carried on the very hands meant to heal. Will we now let arrogance dictate how we use these new instruments? The question of liability is secondary. The first question must always be: Who is still listening to the patient?

July 29, 2026

The question of accountability, when human and machine collaborate in medicine, is presented as novel. It is not. The physician has always relied upon instruments and the testimony of others. A stethoscope does not bear responsibility for a misdiagnosis, nor does a laboratory technician for an incorrect assay. The physician, in integrating these data, makes the final judgment. The burden of proof remains with the one who asserts a conclusion and acts upon it.

The proposed “staging system” for AI’s role in patient care suggests a gradation of responsibility. This implies that responsibility itself is a fluid concept, rather than a fixed point. Either one is accountable for a decision, or one is not. If the AI merely presents information, as a textbook might, then the physician’s accountability is unchanged. If the AI makes the decision, then the question shifts: who programmed the AI, and on what basis? The error, if one occurs, resides in the premise or the execution, both human constructs.

To suggest that accountability can be partitioned is to misunderstand its nature. It is a singular locus. One cannot be partially responsible for a consequence, as one cannot be partially pregnant. The physician either assents to the course of action or dissents. The tools employed do not dilute this fundamental fact. The evidence for a physician’s culpability rests on whether a reasonable practitioner, given the available information and tools, would have acted differently. The AI is merely another tool, albeit a complex one. Its complexity does not alter the logical structure of accountability.