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.

September 25, 2026

In this week’s Duke AI Health Friday Roundup: hedgehogs vs foxes in an era of AI; AI models training AI models; agentic systems devise candidate chemotherapies; flood of AI papers may herald end of current scholarly paradigm; Anthropic claims Claude discovery of enzyme system; test of RAG patient chatbot reveals security holes; AI experts critique recent hype, doomsaying; sugar deprivation in early life may be associated with health benefits later; much more:

AI, STATISTICS & DATA SCIENCE

“404 Media has found multiple contractors hired to improve OpenAI’s models have been fired for using AI to train the AI….Some AI models already exhibit signs of “model collapse,” which is where AI models further trained on AI-generated text can become worse and worse. In this case, some of the people hired to partially stop that happening are themselves using AI-generated responses to train OpenAI’s models.” In a positive ouroboros of irony, 404 Media’s Joseph Cox reports that some of OpenAI’s contractors have been using AI systems to train OpenAI’s AI, despite explicit directions to the contrary.
“The Virtual Biotech comprises as many as 37,000 agents….It uncovered a molecular signal that could help to predict clinical-trial success. And with some human oversight, it identified a promising lung-cancer treatment….But other scientists note that the Virtual Biotech has not been vetted in the crucible of real-world drug discovery, and its predictions were not validated through experiments, let alone clinical trials.” In a news feature for Nature, Ewen Callaway describes how a team of researchers used a “team” of thousands of AI agents to identify candidate drugs for cancer therapy.
“Using only a standard web browser, its built-in developer tools, and a commercial AI assistant available to any consumer (Claude Opus 4.6), we were able to retrieve — without authentication — the operative system prompt, the full RAG pipeline configuration including model and embedder identifiers, the internal address of the vector database, the documents that constituted the knowledge base, and the 1000 most recent patient–chatbot conversations.” A perspective article published in NEJM AI by Madrid-García and colleagues, the authors describe a test of a patient-facing educational chatbot that revealed serious security issues and make a case for subjecting AI tools – in the case, a chatbot using retrieval-augmented generation (RAG) – to the same kinds of rigorous pre-deployment security testing as other kinds of technology.
“Today, we’re sharing early results from one of our first research programs, in which Claude autonomously discovered a novel enzyme system that is associated with an array of DNA repeats, a pattern reminiscent of CRISPR. Although we don’t yet know its function, the system that Claude discovered has a set of characteristics that have only ever been found together in a handful of other systems, all of which are programmable and perform operations like cutting, copying, and pasting DNA.” An article published by Anthropic announces the “autonomous” discovery of an enzyme system by the company’s Claude AI.

BASIC SCIENCE, CLINICAL RESEARCH & PUBLIC HEALTH

“People who had less sugar exposure during the first few years of life were much less likely to develop several different cancers later in life, and they also showed signs of slower biological ageing. We also found that even in adulthood, they continued to consume less sugar and had healthier diets overall.” In an article for The Conversation, economics professor Weilong Zhong describes recent research in which he and his colleagues took advantage of a natural experiment created by postwar sugar rationing in the United Kingdom to examine possible associations between consumption of sugar early in life and later health outcomes.
“Two parallel brain progenitors emerge simultaneously during gastrulation: anterior neural ectoderm (forebrain/midbrain progenitor) and posterior neural ectoderm (hindbrain progenitor). Differentiation of human pluripotent stem cells into anterior or posterior neural ectoderm-like cells revealed these were lineage committed to forebrain/midbrain versus hindbrain fates, respectively.” A research article published in Nature Neuroscience by Jokhai and colleagues presents findings that suggest different regions of the human brain arise from different cellular progenitors.
“All 31 leading cancer centers responding to a new survey reported shortages of at least one cancer drug, and nearly two-thirds said they resorted to mitigation strategies to keep patients on their recommended treatments….But in many cases, those mitigation strategies weren’t enough. Ninety percent of centers reported that treatment plans had to be modified because of drug shortages, creating additional insurance hurdles that may have delayed care.” The University of Minnesota CIDRAP’s Laine Bergeson reports on ongoing shortages of chemotherapy drugs and the downstream consequences for patients and research participants.
“The current deaths and hospital admissions in Pennsylvania are ‘definitely not proportional to what would be expected’, Houry said, and she believes the official numbers are ‘very much an undercount’. Under-testing can make it challenging to do contact tracing and sick people may not realize they need to isolate – both of which can make the spread of measles more likely.” In an article for The Guardian, Melody Scheiber reports on the expanding measles outbreak in Pennsylvania and the difficulty doctors are having in getting a clear understanding of current outcomes related to the outbreak.

COMMUNICATIONS & Policy

“AI tools will certainly bring efficiency — it’s a lot faster and more manageable to read a three-sentence summary of a paper or to have the numbers behind a figure calculated rather than having to derive them yourself. But, like most things AI, this places all of the value on outputs, rather than on processes. The processes of reading, analyzing, reflecting, and theorizing are important to how we learn to think….The output doesn’t matter, but the process of reaching that output does.” In an essay at Scholarly Kitchen, David Crotty frames scholarly reading as falling roughly into two camps – quick-skimming “foxes” and deep-reading “hedgehogs” – and ask what we can expect as AI-generated summaries seem to tilt the scale in favor of the foxes.
“Free access is worthless when the binding constraint is not access but attention, and attention is allocated by brand. To be honest, we are already seeing that trend and AI will only accelerate. We will have open archives that nobody reads and closed reputational cartels that everybody does. That is a worse settlement than the one we are leaving, and it would be dishonest to present it as an equilibrium anyone should welcome.” An essay by Giorgio Gilest explores the challenges imposed by a rapidly expanding volume of publications, many of the partly or wholly developed using AI systems.
“Claims of incipient, dangerous superintelligence are not based in good scientific or engineering practice….It’s worth thinking about why there is so much attention on computer programming and math as fields in which to apply large language models and related technology. Not only are they often elevated as the pinnacle of human intellectual achievement, but they involve problems where answers, once suggested, can be verified.” In a perspective article for MIT Technology Review, Timnit Gebru and Emily M. Bender offer a skeptical dissection of current hype (and related doomsaying) around current AI capabilities.