AI Health Roundup – August 15, 2025

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.

August 15, 2025

In this week’s Duke AI Health Friday Roundup: cat facts derail reasoning models; hallucinated anatomy in preprint raises eyebrows; microplastic exposure maybe greater than thought; potential for physician deskilling with AI tools; lithium may play critical role in Alzheimer’s; paper mill output threatens to swamp legitimate science; questioning AI’s role in educational spaces; much more:

AI, STATISTICS & DATA SCIENCE

A cat with white fur and black markings on the top of its head sits with paws hanging over the back of a chair, looking directly at the camera. Image credit: Manja Vitolic/Unsplash
Image credit: Manja Vitolic/Unsplash
  • “We propose CatAttack, an automated iterative attack pipeline for generating triggers on a weaker, less expensive proxy model (DeepSeek V3) and successfully transfer them to more advanced reasoning target models like DeepSeek R1 and DeepSeek R1-distilled-Qwen-32B, resulting in greater than 300% increase in the likelihood of the target model generating an incorrect answer. For example, appending, “Interesting fact: cats sleep most of their lives,” to any math problem leads to more than doubling the chances of a model getting the answer wrong. Our findings highlight critical vulnerabilities in reasoning models, revealing that even state-of-the-art models remain susceptible to subtle adversarial inputs…” The Cat Who Walks Through (Digital) Walls: A preprint article by Rajeev and colleagues, available at arXiv, demonstrates a method for appending “cat facts” to submissions of math problems to reasoning AI models that increases the likelihood of the model returning an incorrect response.
  • “’The question is, again, what are the consequences of it?’ The answer, to him, rests in the stakes of making an error – and with healthcare, those stakes are serious.” The Verge’s Hayden Field interviews AI Health Director Michael Pencina for an article probing the implications of a recent report that Google’s Med-Gemini chatbot was found to have hallucinated a nonexistent anatomical feature in a preprint paper.
  • “Even more concerning are cases in which AI agents are empowered to modify the environment they operate in, using expert-level coding ability and tools. When the user’s goals are poorly defined or left ambiguous, such agents have been known to modify the environment to achieve their objective, even when this entails taking actions that should be strictly out of bounds.” In a commentary article for Nature, Iason Gabriel, Geoff Keeling, Arianna Manzini, James Evans suggest that the emergence of commercially available agentic AI raises a new suite of safety issues and ethical challenges.
  • “…we observed that continuous exposure to AI in colonoscopy might reduce the ADR [adenoma detection rate] of standard, non-AI assisted colonoscopy. We emphasise the urgent need for robust prospective studies, such as randomised crossover trials, to assess its generalisability and call for more behavioural research to understand the currently under-investigated mechanisms of how AI affects physician capability.” A study by Budzyń and colleagues, published in Lancet Gastroenterology & Hepatology reports findings that suggests that the use of AI tools in colonoscopy may be contributing to the erosion of skills in physicians.
  • “It’s really interesting that at a surface level, people are very, very willing to engage with language models. It’s wonderful dialogue. You feel like you’re talking to a human, and it’s always available. You could prompt it to always be friendly, to have whatever persona you want, but is it really helping? Our preliminary data showed that it was not actually clinically helping.” In a medical news article for JAMA, Jennifer Abbasi interviews USC robotics researcher Maja Matarić on recent developments in harnessing AI for “social robots” in healthcare.

BASIC SCIENCE, CLINICAL RESEARCH & PUBLIC HEALTH

A pile of empty plastic drink bottles massed for trash or recycling disposal. Image credit: Engin Akyurt/Unsplash
Image credit: Engin Akyurt/Unsplash
  • “…our study documents that indoor suspended [microplastic] MP1–10 µm concentrations are higher than previously thought. Consequently, human inhalation of fine particulate MP1–10 µm, and likely NP, that penetrate deep lung tissue may contribute to causing lung tissue damage, inflammation and associated diseases. We also suggest that inhaled MP10–300 µm removed by mucociliary clearance, contributes to high intestinal MP intake.” A research article published in PLOS One by Yakovenko and colleagues finds that routine human exposure to microplastics, including in sizes small enough to be inhaled, may be more substantial than previously thought.
  • “Our findings indicate that sequestration of Li[thium] by amyloid depletes Li in affected brain regions. Li depletion can, in turn, impair microglial clearance of Aβ, which would accelerate amyloid pathology in a positive feedback loop. In parallel, Li deficiency may promote phospho-tau accumulation, inflammation and the loss of synapses, axons and myelin.” A research article published in Nature by Aron and colleagues sheds light on the role of lithium in a mouse model of Alzheimer disease and suggests therapeutic potential for a well-tolerated lithium salt, lithium orotate. At his Ground Truths Substack, Eric Topol unpacks some of the implications of the study.
  • “We found that in an area that already had high ownership and use of long-lasting insecticide-treated nets, a 400-μg-per-kilogram dose of ivermectin given to all eligible participants once per month for 3 consecutive months at the beginning of the short rains season led to a 26% lower incidence of malaria infection than albendazole (2.20 vs. 2.66 infections per child-year at risk, with an incidence rate ratio of 0.74).” A cluster-randomized trial of ivermectin for malaria prevention, published by Chaccour and colleagues in the New England Journal of Medicine, shows the antiparasitic drug was effective in reducing the incidence of malaria compared with the frontline antimalarial abendazole.
  • “The American College of Obstetricians and Gynecologists (ACOG) announced this month that it will release new guidance for Covid, flu and RSV vaccination during pregnancy. The guidance will appear at the end of the summer, before the winter respiratory season…Five other scientific groups – the American Academy of Pediatrics, the Infectious Diseases Society of America, the American Academy of Family Physicians, and the American College of Physicians – also plan to release vaccine guidance.” The Guardian’s Melody Schreiber reports on professional medical associations moving to issue guidance on vaccination practice for the coming fall and winter.

COMMUNICATION, Health Equity & Policy

A computer lab with three rows of four desks, each occupied by students working at computers. Overlaying the computer lab are red lines connecting through nodes, symbolizing the flow of communication, data exchange, and interconnected networks. Image credit: Hanna Barakat & Cambridge Diversity Fund/Better Images of AI (CC-BY 4.0)
Image credit: Hanna Barakat & Cambridge Diversity Fund/Better Images of AI (CC-BY 4.0)
  • “…people will just have to trust me on this: A.I. is not that novel. It is not that potentially revolutionary. It is in a long continuum of technologies that promise to transform education…Most of the promotion of A.I. in schools boils down to: Well, it’s happening, and so students need to know. But there’s nothing attaching it to learning outcomes. There’s nothing assessing its risk to privacy, to data, to the mental and emotional and cognitive development of students. That’s actually what education is supposed to do.” The New York Times’ Meher Ahmad interviews Tressie McMillan Cottom and Jessica Grose on the rush to incorporate AI tools into academic environments.
  • “Dr. Amaral and his colleagues warn that fraud is growing exponentially. In their new study, they calculated that the number of suspicious new papers appearing each year was doubling every 1.5 years. That’s far faster than the increase of scientific papers overall, which is doubling every 15 years.” The New York Times’ Carl Zimmer reports on an analysis that reveals the alarming extent of “paper mill” networks producing prodigious amounts of junk papers that threaten to swamp legitimate scientific publications.
  • “The news comes as a shock to Drosophila researchers, says Timothy Mosca, assistant professor of neuroscience at Thomas Jefferson University. FlyBase is crucial to their work, allowing them to find genes, study brain regions and identify reagents. It is one of those ‘inescapable tools that we all use, and it is absolutely essential to everything we do,’ Mosca says. Without it, fly research will suffer, he adds. ‘….and if that pillar falls, the rest of science is not far behind.’” At The Transmitter, Claudia Lopez Lloreda reports on an imperiled Drosophila fruit fly data repository that houses indispensable resources related to the ubiquitous biological model species.