Doctors Must Keep the Weight of Responsibility That AI Cannot Carry

Artificial intelligence is moving through medicine at a pace that has unsettled even the most seasoned clinicians. Systems can now read electronic health records, synthesize laboratory data and written case narratives, and arrive at diagnoses that match or in some cases exceed the performance of human physicians. That capability has reignited a familiar and anxious question: if a machine can diagnose as well as a doctor, what exactly is the doctor for? Two prominent physician-researchers believe they have an answer, and it is one that resists both the hype of AI enthusiasts and the fatalism of those who fear wholesale replacement.

Andrew S. Parsons, MD, MPH, of the University of Virginia School of Medicine and UVA Health, and Adam Rodman, MD, MPH, of Harvard Medical School, have published a commentary in the Annals of Internal Medicine arguing that the enduring core of medicine is not diagnosis at all. It is accountability. Their central claim is deceptively simple: even as AI absorbs more of the cognitive labor of identifying disease, the human clinician must remain the party who reasons through what to do next, works with the patient to make difficult choices, and answers personally for the outcome. That responsibility, they contend, cannot be offshored to an algorithm, no matter how accurate its outputs become.

The distinction the authors draw is between two very different kinds of clinical judgment. The first is the diagnostic task: given a set of findings, what condition best explains them? This is a problem with, in principle, a correct answer, and it is precisely the kind of pattern-recognition challenge at which machine learning excels. The second kind of judgment is managerial and deeply personal: which treatment, which tests, which follow-up plan actually fits this particular patient, with their goals, their comorbidities, their social circumstances, and their tolerance for risk? That question, Parsons and Rodman argue, cannot be reduced to a single correct answer, because the answer depends on whose life is being lived and whose values are being weighed.

The evidence base that guides clinical decisions makes this problem sharper, not softer. Guidelines and randomized trials, the authors note, skew toward the populations studied in those trials. Real patients frequently fall outside those norms. They carry complex combinations of chronic conditions, they live in social circumstances the trial never anticipated, or their personal goals simply differ from the end points a trial was designed to measure. In such cases, deviating from the technically correct answer is often good medicine. But a deviation is a bet, and someone must answer for it. The commentary insists that the person who answers must be a human being with a stake in the result, not a system with nothing to lose.

This framing gives the clinician’s role a precise technical definition rather than a sentimental one. Parsons and Rodman write that the clinician’s enduring function is accountability: owning the risk of acting under uncertainty and answering for the decision that follows. Crucially, they add a condition that separates genuine accountability from a rubber-stamp ritual. That ownership is real only when the clinician still reasons about the decision and holds a stake in its outcome. A doctor who merely forwards an AI recommendation to a patient without independent reasoning has not accepted responsibility; they have laundered it. The commentary’s implicit warning is that medicine must not slide into that posture as automation deepens.

Why does human ownership change the quality of the decision itself? The authors offer a behavioral argument grounded in the texture of ongoing clinical relationships. A clinician who knows they will answer for the result weighs options differently than a system with nothing to lose. A clinician who expects to see the patient again also considers which choices will preserve the relationship if things go badly. These are not soft considerations; they are consequences that feed back into decision-making in real time. An algorithm can optimize for a probabilistic outcome across a population, but it does not have to face the patient, the family, or the complication at the bedside next week. The asymmetry of stakes, the authors suggest, is not a defect to be engineered away but a structural feature of moral agency in medicine.

None of this amounts to a call for resistance. Parsons and Rodman explicitly urge physicians to embrace the technology. AI will keep improving the information that feeds clinical decisions, and better information is an unambiguous good. The diagnostic accuracy of these systems is already impressive and will likely continue to climb, and clinicians who refuse to use them may deprive patients of genuinely useful insight. The commentary’s target is a specific temptation: the idea that because AI can answer diagnostic questions well, the entire encounter of care can be delegated to it. The authors’ answer is that diagnosis was never the whole of medicine, and the part that remains is the part that matters most.

The timing of the argument is significant. Debates about AI replacing doctors have intensified as large language models demonstrate the ability to work through written clinical cases and propose plausible differential diagnoses. Some commentators have treated diagnostic parity as the finish line, imagining a future in which the physician is a vestigial supervisor. Parsons and Rodman reframe the question entirely. If the irreplaceable human contribution is the assumption of responsibility under uncertainty, then the relevant metric for automation is not accuracy on answerable questions but capacity for answerability itself, and no current or foreseeable system can hold that capacity. The patient, after all, is the one who must live with the consequences, and the clinician is the one who must stand beside them when those consequences arrive.

The commentary also carries a quiet warning about how the profession organizes itself in the years ahead. If health systems treat AI as a way to strip clinical roles down to their most easily automated components, they risk producing clinicians who no longer reason deeply about decisions, which by the authors’ own criterion would dissolve the very accountability that justifies the clinician’s existence. Preserving the human role therefore requires deliberate design: workflows in which AI informs but does not finalize decisions, training that keeps physicians engaged in the hard judgment of individualizing care, and institutional structures that keep a named human answerable for every consequential choice. The alternative, a medicine of frictionless delegation, would be cheaper and faster in some respects, but it would leave no one holding the risk.

Parsons has no financial interest in the work; Rodman has done unrelated work for Google. The commentary, published in the Annals of Internal Medicine, arrives as health systems worldwide are piloting AI tools in documentation, triage, and diagnosis, and it offers those institutions a principle for drawing the line. Welcome the machine’s growing accuracy, the authors conclude, but do not hand it the weight. Medicine of tomorrow will still need someone to choose which risks a particular patient should accept, to make the hard choices alongside them, and to shoulder the responsibility for what follows. That is not a task AI has failed at; it is a task that, by its nature, only a human can perform.

Subject of Research: The role of human accountability in clinical decision-making as artificial intelligence takes on diagnostic tasks in medicine

Article Title: The crucial responsibilities doctors must never offshore to AI

Article References: The crucial responsibilities doctors must never offshore to AI. (n.d.). Original publication

Image Credits: AI Generated

DOI: Not provided

Keywords: artificial intelligence, clinical decision-making, medical accountability, diagnosis, physician role, Annals of Internal Medicine, patient care, health policy, medical ethics, clinical guidelines, healthcare automation, UVA Health