AI Health
Friday Roundup
The AI Health Friday Roundup highlights the week’s news and publications related to artificial intelligence, data science, public health, and clinical research.
April 24, 2026
In this week’s Duke AI Health Friday Roundup: favoring owls and other signs of AI misalignment; sifting through health queries to Copilot; gamification for emergency triage training; LLMs, EHR notes, and accountability; progress with JEPA world models; the “mixed bag” of the current peptide moment; AI ushers in new era of risks for health system data and infrastructure; much more:
AI, STATISTICS & DATA SCIENCE
- “In our main experiments, a ‘teacher’ model with some trait T (such as disproportionately generating responses favouring owls or showing broad misaligned behaviour) generates datasets consisting solely of number sequences. Remarkably, a ‘student’ model trained on these data learns T, even when references to T are rigorously removed. More realistically, we observe the same effect when the teacher generates math reasoning traces or code. The effect occurs only when the teacher and student have the same (or behaviourally matched) base models.” A research article published in Nature by Cloud and colleagues shows evidence for “subliminal learning” that can transmit preferences during the process of model distillation, in which one model is used to train another.
- “While LLMs reduce administrative burden, their integration in clinical workflows introduces the risk of blending AI- and human-generated content within EHRs. This perspective reviews technological and policy solutions to ensure traceability of AI-generated content in EHRs to preserve clinical integrity.” A preprint article by Nargessi and colleagues, accepted and in press at NPJ Digital Medicine, examines complications posed by the creep of AI-authored text into patient electronic health records, and the challenges that creates for establishing accountability.
- “With ~15M parameters trainable on a single GPU in a few hours, LeWM plans up to 48x faster than foundation-model-based world models while remaining competitive across diverse 2D and 3D control tasks. Beyond control, we show that LeWM’s latent space encodes meaningful physical structure through probing of physical quantities. Surprise evaluation confirms that the model reliably detects physically implausible events.” A preprint by Maes and colleagues, available from arXiv, unpacks an evaluation of an new approach for training world models using joint embedding predictive architectures.
- “Going forward, the medical AI field must develop a consistent framework to connect claims of clinical value of an AI tool to the appropriate type of evidence needed to support those claims. For example, claims of analytic performance should require robust validation in the intended setting and population, whereas claims of clinical actionability should require evidence that outputs are interpretable and can support reasonable decisions.” A statement by the editors of Nature Medicine calls for more rigorous and apposite standards for determining whether a clinical AI application actually improves patient care.
- “The prevalence of queries related to finding providers, understanding insurance and completing paperwork reveals that a meaningful fraction of health AI use addresses the complexity of healthcare systems rather than health itself. Users are asking AI to help them do things that should, in principle, be straightforward: find a doctor, book an appointment and understand what their insurance covers.” A research article by Costa-Gomes and colleagues, published in Nature Health, parses a month’s worth of health-related queries to Microsoft Copilot.
BASIC SCIENCE, CLINICAL RESEARCH & PUBLIC HEALTH
- “How about the science? It’s the biggest mixed bag you ever saw. There’s no doubt at all that there are some extremely biologically active peptides out there…that’s the first point: there are indeed a whole range of physical and medical effects to be found in these things….Unfortunately, point two is that we barely have any of these effects worked out – at least not to the degree that you would want before you start injecting them into your leg.” In an article at his In the Pipeline blog, Derek Lowe tackles the current fad for “peptide” supplements for a dizzying array of health-related (or health-adjacent) applications.
- “This analysis found that launch of the 988 Suicide and Crisis Lifeline was associated with significant reductions in suicide mortality among adolescents and young adults nationally. Observed reductions were larger in states with the highest uptake of 988 Lifeline services.” A research letter published in JAMA by Patel and colleagues analyzes data from an observational study that examines adolescent suicide rates since the creation of the “988” suicide prevention hotline.
- “With the launch of UT Dell Medical Center, however, Dr. Claudia Lucchinetti sees a rare opportunity: instead of retroactively applying new technologies to old hospital infrastructure, she said they can integrate them from the start. They will also collaborate with the University of Texas MD Anderson Cancer Center in Houston to offer top specialists for those with complex conditions.” The Associated Press is reporting on a gift of $750 million from Michael and Susan Dell to the University of Texas to establish an AI-first medical school.
- “In this randomized clinical trial, we found that exposure to a customized, theory-based serious game reduced undertriage of patients by physicians working in the EDs of nontrauma centers in the US without increasing overtriage of patients by physicians.” A research paper published in JAMA by Mohan and colleagues describes results from a randomize trial that used gamification to address adherence to triage guidelines in emergency physicians.
COMMUNICATIONS & Policy
- “The same capabilities being celebrated for drug discovery are now powerful enough to find and weaponize software vulnerabilities at machine speed, and health care’s defenses weren’t built for that pace. And while health care has been racing to find cures with artificial intelligence, nation-states have been in an arms race to wield power over each other. This adversarial landscape is compounded by a race between Silicon Valley and health systems to compete, often with each other.” In an opinion article for STAT News, The Light Collective’s Andrea Downing sounds an alarm on the vulnerability of healthcare systems’ information infrastructures in the face of AI-powered hacking attacks.
- “The study, conducted by C.D.C. scientists, calculated the effectiveness of Covid shots by looking at the vaccination status of people who had sought care at hospitals and emergency rooms. It found that vaccination cut the likelihood of emergency visits due to Covid by 50 percent and of hospitalizations by 55 percent, according to a summary of the study viewed by The New York Times.” The New York Times’ Apoorva Mandavilli reports that CDC officials have halted the expected publication of a study by CDC researchers that examined health outcomes associated with receiving vaccination for COVID.
- “Polling mainstays like Qualtrics and newer firms like Synthetic Users are proffering ‘silicon’ respondents—large language models (LLMs) that will pretend to be lots of different people and answer questions based on how the models predict they would respond.” At Mother Jones, Henry Carnell reports on the rise of real-world use of AI respondents in public polling.
- “AI adoption in higher education is not just a matter of technical exposure, but also requires other factors, such as psychological and structural readiness. Given the growing influence of AI, students require both confidence and institutional support to utilize AI-driven tools effectively. The findings underscore the need for higher education institutions to actively develop AI literacy programs and foster a culture that promotes the responsible use of AI.” A research article by Fute and colleagues, in press at the journal Humanities and Social Sciences Communications, examines conditions needed for deliberate adoption of AI tools in undergraduate and graduate settings.
