Menu
Close
.svg)
As artificial intelligence becomes increasingly adept at clinical reasoning, pharma faces a more consequential question than how many HCPs use it: what happens when human and machine judgment collide?
This hypothetical, seemingly future-forward scenario has undoubtedly happened already. Imagine an oncologist reviewing the case of a patient whose disease has progressed despite treatment. Drawing upon years of clinical experience, the physician herself favors a familiar therapeutic approach, one whose benefits and limitations are well understood. An AI system, having examined the patient's history, molecular characteristics and emergent clinical evidence, recommends something different. Its reasoning is cogent, its supporting evidence persuasive, and its conclusion at odds with the physician's instincts. What happens next?
Does the oncologist reconsider the original recommendation, interrogate the AI's reasoning or simply dismiss its conclusions? Might a younger physician respond differently from a seasoned specialist? And what happens if or when the algorithm is wrong, but expresses its judgment with a confidence and authority that belies its uncertainty? These questions are no longer fanciful thought experiments. They are becoming increasingly pertinent to the practice of medicine, and to the pharma industry's understanding of how treatment decisions are made.
A remarkable study published last week in Science offers a glimpse of what lies ahead. Here is the link: https://www.science.org/doi/10.1126/science.aeh7930?referrer=https%3A%2F%2Fwww.google.com%2F Researchers at Microsoft developed an AI system capable of approaching diagnosis sequentially, much as physicians do, gathering information, ordering further tests, and continually revising its hypotheses as new evidence is revealed. Rather than merely supplying an answer to a neatly packaged clinical question, the system tries to navigate the uncertainties of diagnosis itself.
The results were arresting. Across challenging cases drawn from the New England Journal of Medicine, the system achieved diagnostic accuracy of 83.2%, compared with 36.1% among generalist physicians permitted to consult reference materials. It also performed strongly on emergency department cases, with favorable accuracy and testing-cost trade-offs under several conditions. These were controlled, text-based evaluations, not demonstrations of superiority in everyday clinical practice, and the researchers did not assess treatment decisions or patient outcomes. Nevertheless, the findings suggest that AI is acquiring capabilities that extend well beyond straightforward medical information retrieval.
Meanwhile, the American Medical Association's recent 2026 survey found that 81% of physicians now use AI professionally, more than twice the proportion recorded in 2023. Yet only 17% reported using it for assistive diagnosis, with considerably more relying on the technology for research summaries and administrative tasks. We in pharma therefore find ourselves in an intriguing position: widespread adoption of AI alongside a much less developed understanding of its influence on clinical judgment.
This is where behavioral science becomes especially interesting. Physicians, like everyone else, are susceptible to anchoring, confirmation bias and the influence of established mental models, aka habits. Clinical experience provides invaluable pattern recognition, but can also make familiar explanations disproportionately appealing. An AI recommendation that challenges an initial diagnosis might therefore serve as a useful corrective, prompting reconsideration of possibilities that would otherwise receive scant attention.
Yet the opposite danger is equally plausible. Automation bias describes our propensity to accord undue weight to recommendations generated by automated systems, sometimes at the expense of contradictory evidence. A physician confronted with a confident, elaborately reasoned AI recommendation may defer to it even when clinical intuition suggests otherwise. Conversely, algorithm aversion might encourage physicians to discount sound advice after encountering an earlier error. Neither uncritical acceptance nor reflexive rejection constitutes particularly good medicine.
Recent research adds another wrinkle. A June 2026 preprint examining physicians' interactions with agentic AI found that more transparent, stepwise reasoning increased trust and reliance, although physicians sometimes placed excessive confidence in incorrect outputs. The findings suggest that making an algorithm's reasoning visible may encourage acceptance without necessarily ensuring that such acceptance is warranted. The challenge, therefore, is not simply to cultivate trust in AI, but to calibrate it appropriately.
For those of us working in I&A supporting pharma decision-making, these developments raise questions that conventional prescribing research is ill-equipped to answer. We routinely ask about the clinical evidence physicians value, the attributes they associate with competing therapies, and the circumstances under which they might switch treatments. But how much do we know about the AI systems they consult, the questions they pose, or the recommendations that subsequently influence their judgments?
Consider the commercial implications if AI begins identifying previously overlooked diagnoses, challenging established treatment approaches, or drawing attention to therapeutic alternatives that physicians might not otherwise consider. Equally consequential is the possibility that algorithms reinforce existing prescribing conventions, privileging familiar treatments or widely represented evidence over newer approaches. These are competing hypotheses, and we need empirical research to distinguish between them.
I believe pharmaceutical marketing research must begin examining physician-AI interactions directly, particularly at moments of clinical uncertainty. We should investigate when physicians seek algorithmic assistance, how they formulate their questions, what happens when recommendations conflict with their initial judgments, and whether those encounters ultimately alter diagnosis, treatment selection or patient management. Such research must extend beyond retrospective accounts of AI usage to the actual processes through which human and machine reasoning converge, diverge and influence decisions.
The most consequential development in medical AI may not be that machines are becoming increasingly capable of reaching clinical conclusions. It may be that physicians will increasingly arrive at their own conclusions through a dialogue with those machines, sometimes finding their judgment reinforced, sometimes challenged...and occasionally displaced.
For an industry whose fortunes depend upon understanding the intricacies of clinical decision-making, that evolving dialogue deserves considerably closer scrutiny. Knowing that a physician uses AI tells us remarkably little about its influence. What matters is what the physician thinks, does and ultimately decides differently because of it.
