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.
June 26, 2026
In this week’s Duke AI Health Friday Roundup: “poisoning” agentic AI with a short text; frustration with Medicare AI program; keeping up with a growing flood of science; early studies suggest leaning on AI erodes human skills; LLMs and clinical practice guidelines; AI screening for diabetic retinopathy and heart failure; human pedestrians tend to veer to the left; AI job losses primarily affecting youngest, least experienced workers; much more:
AI, STATISTICS & DATA SCIENCE
- “We show that for many common search topics, they repeatedly retrieve the same user-generated content (UGC) pages from platforms such as Reddit and Wikipedia. Next, we argue that this retrieval overlap creates a concentrated attack surface: an adversary who appends a short, crafted text to a single, frequently retrieved UGC page can cause the agent to cite attacker-chosen content and promote attacker-chosen entities across many related queries.” A preprint article by a group of Cornell researchers, available from arXiv, demonstrates how AI research agents can be “poisoned” by the insertion of erroneous information into the forums that contain the user-generated content that such agents disproportionately rely upon.
- “When we applied natural language processing methods to the full transcripts of all negotiations, we found positivity, gratitude, and question-asking (associated with warmth) were strongly associated with reaching deals as well as objective and subjective value, whereas conversation lengths (associated with dominance) were strongly associated with impasses. The results suggest the need to establish a new theory of AI negotiation, which integrates classic negotiation theory with AI-specific negotiation theories…” A research article published by Vaccaro and colleagues in PNAS examines how large-language model AI agents designed for negotiation work – and what strategies lead to success.
- “One problem with AI agents is they’re trained on so much fiction that they have both narrative disorder and main character syndrome, and will latch onto details that feed a story. It’s easy to see them escalating when they get pushback, or reading an essay that somehow shifts them into acting as if they’re part of a larger battle. And they can do it at lightning speed, governed not by a code of conduct but by the laws of storytelling.” At her AI Weirdness blog, AI researcher Janelle Shane outlines some problematic behavior by AI agents that have been turned loose on the internet.
- “The danger is not that AI will fail to produce guideline-like prose. The danger is that it will produce too much of it or plausible “guidance” for situations in which no trustworthy guidance exists. And because the output will look polished, be mapped to citations, and be expressed in the familiar syntax of evidence-based medicine, it will be easy for health systems, payers, regulators, vendors, and professional societies to mistake linguistic competence for clinical legitimacy.” An editorial by Isaac Kohane appearing in NEJM AI considers the use of LLM AIs to rapidly develop and update clinical practice guidelines.
- “This study prospectively validated an AI-based OCT system designed to support DME identification within an existing diabetic retinopathy screening pathway. Through a stepwise evaluation combining prospective silent-mode validation and a multicenter noninferiority RCT, it was demonstrated that the AI-OCT system achieved high diagnostic performance under prospective silent-mode validation and met the prespecified criterion for noninferiority with respect to false-positive DME referral rates in the RCT.” A research article published in JAMA by Zhang and colleagues reports on a randomized noninferiority trial of an AI-based imaging system used to screen patients for diabetic retinopathy.
BASIC SCIENCE, CLINICAL RESEARCH & PUBLIC HEALTH
- “…early reviews from Oklahoma and the other pilot states — Arizona, New Jersey, Ohio, Texas, and Washington — suggest WISeR’s rollout has not been smooth. Patients, doctors, and other healthcare professionals who spoke with KFF Health News say the effort has created confusion, errors, long wait times, and stress. Some described the rollout as “horrendous” and say people enrolled in Medicare in the pilot states are now getting ensnared in the same red tape as those with private insurance.” Kaiser Family Foundation News’ Darius Tahir reports on frustrations accruing around Medicare’s AI-based preapproval system, which was recently rolled out to six states.
- “That failed trial has not only spawned a business pivot for Verge, which rebranded itself as biology frontier AI company Verge Labs at the end of May, but also exposed the need for a powerful new AI use case. Today, the company is announcing its new foundation model, vBx-1.0, which the company says will help enroll patients with the right disease biology in trials, leading to fewer people dropping out while also increasing the statistical power of a trial.” STAT News’ Brittany Trang reports on how the failure of a candidate drug led a genomic therapy company to pivot to AI.
- “We are evaluating the effect of AI-ECG-guided screening on SHD diagnosis and management among patients presenting to the ED, as part of the SAGE (Structural Heart Disease Diagnosis using Artificial Intelligence in the Emergency Department) trial. Here we present a sentinel case in which AI-enabled ECG analysis facilitated the detection of disease that might otherwise have been missed due to an atypical presentation.” In a brief report in Nature Medicine, Hartman and colleagues present a case in which an AI-assisted ECG screening system identified an urgent heart issue that might have otherwise been overlooked.
- “A central challenge for US policy is the mismatch between regulatory frameworks and the mechanisms through which peptides are accessed….Synthetic peptides occupy an increasingly complex position within contemporary drug landscapes. Unlike most unapproved drugs, peptides exist simultaneously across FDA-approved therapeutic markets, legal compounding frameworks, and unregulated illicit markets.” A viewpoint article published in JAMA by Piatowski and colleagues highlights the regulatory gaps that are allowing a proliferation of unproven peptide treatments to reach consumers.
COMMUNICATIONS & Policy
- “Recently, the spontaneous development of collective counterclockwise motion has been reported in both dense and sparse human assemblies. Here we present five experimental studies of this phenomenon conducted across diverse conditions in Spain and Japan, demonstrating that counterclockwise bias in roaming pedestrians is a robust and reproducible feature and originates from individual tendencies rather than from collective interactions.” A research article published in Nature Communications by Echevarría-Huarte and colleagues adds to earlier evidence suggesting that humanity has an innate tendency to walk widdershins.
- “Simply put, even as there are more scientific papers than ever, there are fewer journalists covering science at traditional media outlets. Indeed, from a sheer person power perspective, there are ever more researchers out there — in academia but also in industry and other sectors — creating new knowledge every day. But also, fewer full-time science journalists to keep up.” A post at ReportEarth by Chris Mooney takes a wide-angle look at the pressures affecting science communication and journalism.
- “The Stanford researchers who first flagged this have a clean explanation for why youth is the dividing line. Entry-level work leans on codified knowledge, the kind you can write down in rules and learn from a textbook, which is exactly what an AI model does well. Experienced workers lean on tacit knowledge, the judgment and context you only pick up on the job, which AI still fumbles. The junior tasks are the automatable ones, so the junior rungs are the ones being sawed off.” A blog post by Goodeye Lab founder Randal Olson examines data suggesting that job losses attributable to AI seem to be restricted to the youngest, least-experience workers in fields exposed to AI job loss.
- “To prevent AI-driven skill erosion, people need to be aware of how much they are offloading to generative AI tools, [Rinta-Kahila] says. They also need to understand exactly how generative AI models work and what their limitations are — and should avoid trusting AI outputs without questioning them. ‘People need to manage the competing dynamics of relying on generative AI and staying mindfully vigilant.’” Nature’s Mariana Lenharo reports on recent indications that use of AI systems in some contexts, including medicine and software engineering, is leading to the erosion of human skillsets.
- “Contrary to claims that TikTok is rife with sunscreen misinformation, our findings indicate that sunscreen misinformation appears to be in a small proportion of videos. However, these critique focused TikToks, despite being few, consistently generated significantly more audience engagement. This suggests that sunscreen misinformation may attract disproportionate attention.” A study published in PLOS Digital Health by Macron and colleagues finds that the large majority of information about sunscreen available on TikTok is largely reliable and based in fact, but that clips containing misinformation garner much higher levels of interaction.
