US Medical Residencies Are Flying Blind on AI Training, National Survey Finds

Artificial intelligence has slipped into the daily workflow of American medicine faster than almost any technology in recent memory, and a new national survey suggests that the training system responsible for producing the next generation of physicians has not kept pace. In a research report published in the Journal of General Internal Medicine, a team of medical educators led by Eliana Bonifacino of MedStar Washington Hospital Center describes the results of a nationwide survey examining how internal medicine residency programs in the United States approach education about artificial intelligence. The study, conducted through the annual survey of the Association of Program Directors in Internal Medicine, offers one of the most systematic snapshots to date of how residency leadership perceives, implements, and struggles with AI-related teaching at the graduate medical education level.

The survey’s methodology reflects the kind of rigor that national workforce studies in medical education typically demand. Program directors, the physicians who bear ultimate responsibility for the curriculum and competency of every resident in their programs, were invited to complete a questionnaire developed by the authors and refined through both pretesting and live pilot testing with members of the Alliance for Academic Internal Medicine. The survey team, which included Verity Schaye of NYU Grossman School of Medicine, Andrew Olson of the University of Minnesota, Marygrace Zetkulic of Hackensack University Medical Center, Michael Kisielewski of the Alliance for Academic Internal Medicine, and Casey McQuade of the University of Pittsburgh, built statistical weighting into the analysis plan and explicitly estimated nonresponse bias, techniques that help ensure the reported results represent the full national landscape of internal medicine training programs rather than just the subset of directors motivated enough to reply. The study was reviewed by an independent institutional review board and deemed exempt, and participants provided informed consent through the survey itself.

Why does formal AI education in residency matter so much? The answer lies in how quickly large language models and related tools have penetrated clinical environments. Unlike previous waves of clinical technology, such as electronic health records or computerized order entry, generative AI systems are not deployed through a single institutional gateway. Residents can open a chatbot on a personal phone between patient encounters, paste in a de-identified or, worse, an identifiable clinical vignette, and receive fluent, confident, and occasionally fabricated medical advice within seconds. Surveys of internal medicine residents published in recent years have documented substantial self-directed use of these tools alongside ambivalent attitudes, with residents reporting both enthusiasm for the technology’s potential and unease about its reliability. That combination of widespread informal use and limited formal instruction is precisely the gap the new study set out to measure.

The technical risks of unsupervised AI use are not hypothetical. Research on large language model hallucinations, the phenomenon in which a model generates plausible-sounding but factually false content, has documented user-reported fabrication across consumer AI applications, and the medical domain is particularly unforgiving of such errors because the consequences fall on patients. A hallucinated drug interaction, a nonexistent guideline citation, or a subtly wrong diagnostic suggestion can propagate into clinical documentation and decision-making before any human catches it. At the same time, evidence is accumulating that these tools can meaningfully influence clinical reasoning in both directions. A randomized clinical trial published in JAMA Network Open in 2024 found that access to a large language model affected physicians’ diagnostic reasoning in measurable ways, and subsequent work on clinician-AI collaborative workflows has explored how the technology might function less as a passive tool and more as an active teammate in the diagnostic process. Both the promise and the peril argue for structured education, because untrained clinicians are poorly positioned to calibrate when to trust, when to verify, and when to ignore an AI output.

Against this backdrop, the survey’s central finding carries a sobering implication: formal education on artificial intelligence within US internal medicine residencies remains inconsistent and underdeveloped, leaving programs to improvise curricula in an area where few feel equipped to teach. Program directors reported on whether their programs offered dedicated instruction on AI, what topics that instruction covered, and what barriers stood in the way of doing more. The picture that emerges is one of a training enterprise aware that it must act but constrained by a shortage of faculty expertise, an absence of validated curricular materials, competing demands on limited educational time, and the sheer speed at which the underlying technology changes. A curriculum module written eighteen months ago may already reference tools or capabilities that have been superseded, a challenge that conventional medical education, accustomed to updating content on the scale of years rather than months, is structurally unprepared to meet.

The study did not emerge in an institutional vacuum. A companion national survey examining artificial intelligence in the internal medicine clerkship, published by an overlapping author group in the same journal, documented similar dynamics at the undergraduate medical education level, suggesting that the gap begins before residency and persists through it. Meanwhile, accrediting and standard-setting bodies have begun to respond. The Accreditation Council for Continuing Medical Education issued guidance in 2026 on the use of artificial intelligence in accredited continuing education, acknowledging that the technology now touches how physicians learn as well as how they practice. Scholars of medical education, including Marc Triola and Rodman’s collaborators writing in Academic Medicine, have argued that integrating generative AI into medical training requires coordinated attention to curriculum, policy, and governance, rather than leaving individual instructors to improvise. The new residency survey provides the empirical baseline against which such proposals can be judged.

What would adequate AI education for internal medicine residents actually look like? The literature points toward several core components. Residents need a working technical literacy: an understanding of what large language models are, how they are trained on vast text corpora, why they generate probabilistic rather than verified outputs, and why their fluency can mask unreliability. They need prompt-level skills, the ability to structure a query so that a model’s strengths are exploited and its weaknesses exposed. They need verification habits, including the discipline of checking AI-generated citations and clinical claims against primary sources before acting on them. They need to understand privacy and data governance, because entering patient information into unapproved consumer tools can violate institutional policy and health privacy law. And they need to understand bias, since models trained on historical medical data can reproduce and amplify documented inequities in care. None of these topics requires a residency program to teach computer science, but all of them require faculty who are themselves conversant enough to lead the discussion.

The barrier most frequently implicated in the survey context, faculty expertise, illustrates a structural irony of the moment. The physicians most fluent in AI are often early in their careers or embedded in technical fields, while the educators responsible for residency curriculum were trained in an era when the dominant computational tools of medicine were databases and search engines. Closing that gap will require deliberate faculty development, partnerships between clinical departments and informatics or data science units, and shared curricular resources that individual programs can adopt rather than each rebuilding from scratch. The authors’ decision to publish their survey summary results and methodology openly, including the statistical weighting plan and nonresponse bias estimates, reflects a broader push within the Alliance for Academic Internal Medicine to make national educational research transparent and reusable.

The stakes extend well beyond the residency classroom. Today’s residents will practice for decades in environments where AI is woven into documentation, diagnostic support, triage, and patient communication. If their only training comes from trial and error with consumer chatbots, the profession will have abdicated one of its core responsibilities: ensuring that every physician who holds itself out as competent can use the tools of contemporary practice safely and critically. The survey also carries a message for patients, who increasingly encounter AI-generated content in their care, from portal messages to after-visit summaries, often without knowing it. A physician workforce trained to interrogate those outputs is a direct patient safety safeguard.

What happens next will depend on whether the survey’s findings catalyze action at the level of accreditation, specialty societies, and individual programs. The infrastructure already exists: national program director networks, annual surveys that can track progress over time, and a growing evidence base on how AI affects clinical reasoning. The Journal of General Internal Medicine study, received in June 2026 and published in September, arrives at a moment when the question is no longer whether internal medicine should teach about artificial intelligence but how quickly programs can build the expertise to do it well. The survey’s most important contribution may be its candor: it documents, with national scope and methodological care, exactly how far the training system still has to go, and it gives educators, accreditors, and policymakers a shared factual baseline from which to start closing the gap.

Subject of Research: AI education in US internal medicine residency programs

Article Title: Education on Artificial Intelligence in US Internal Medicine Residencies: Results of a National Survey

Article References: Education on Artificial Intelligence in US Internal Medicine Residencies: Results of a National Survey. (n.d.). https://doi.org/10.1007/s11606-026-10811-x

Image Credits: AI Generated

DOI: 10.1007/s11606-026-10811-x

Keywords: artificial intelligence, medical education, internal medicine residency, large language models, program directors, national survey, graduate medical education, hallucinations, clinical reasoning, curriculum, generative AI, patient safety