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.
July 10, 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
- “Overall, AI tools showed strong potential to enhance efficiency and reduce manual workload and costs. Studies reported high accuracy in pre-screening and exclusion of ineligible patients, reliable precision in extracting and defining eligibility criteria, and strong performance in classifying criteria with high sensitivity, precision, and accuracy. Cohort and site identification showed moderate discriminative performance, highlighting areas for further model refinement.” A research article published in the Journal of Clinical and Translational Science by Yin and colleagues presents a scoping review and analysis of the use of AI in recruiting and retaining participants for clinical trials.
- “Ford had been increasingly relying on AI-driven inspection systems to streamline production and address quality control issues, however the firm acknowledged that AI lacked the nuanced judgement when it came to complex problems. After rehiring experienced engineers, Ford experienced a marked improvement in its quality standards.” The Independent’s Anthony Cuthbertson reports on the reversal of layoffs at Ford after an emphasis on AI over human engineering expertise failed to yield the desired results.
- “Over the past few years, some hospital networks have outsourced parts of their core workforces to AI-powered labor platforms like ShiftMed, CareRev and Clipboard Health….Nurses on these platforms report working for lower wages, competing for shifts and having to bring equipment that would normally be paid for by an employer, like stethoscopes and thermometers.” At The Guardian, Arielle Pardes explores the potential for AI to accelerate trends toward the “gigification” of employment and increasing precarity in the workforce.
- “Denial, override, and escalation rates can be tracked continuously rather than audited episodically. The LLMs are rightly criticized as black boxes because their internal logic is not transparent enough to satisfy human intuitions. But in adversarial administrative settings, the operative contrast is no longer transparent humans vs opaque machines. It is humans, whose decisions may be understandable one at a time but expensive to characterize at scale, against AI systems that may remain internally opaque yet become benchmarkable when properly logged.” In a perspective for JAMA, Isaac Kohane outlines how “black box” AI systems could paradoxically be used to make payer coverage decisions more legible.
- “Overall, the LLMs were from 13 to 17 times more likely to produce a negative story for a character with health conditions than for a healthy one. The models were less biased than humans: When the researchers presented the same scenarios to 399 people, they were up to 23 times more likely to write a negative story for characters with health conditions. Still, the results suggest LLMs, while filtering out some stigma, were mimicking human biases embedded in their training materials.” Science’s Laura Martín Agudelo reports on recent research showing that output from widely used chatbots can reinforce stigma around health issues.
BASIC SCIENCE, CLINICAL RESEARCH & PUBLIC HEALTH
- “The company’s new product will not achieve such lofty goals immediately, and Anthropic doesn’t pretend it will. The Claude Science tool uses the same reasoning model as other Claude products, but it is customized so that scientists can import all sorts of different types of data sources — everything from DNA sequences to cell images to microscopy data — and allow the large language model to both help with rote tasks and provide ideas.” STAT News’ Matthew Herper and Brittany Trang report on the release of Claude Science, a version of the Anthropic LLM that’s specifically trained for use in the life sciences.
- “Though Kauderer-Abrams did not disclose what indications Anthropic is pursuing, another executive, Jonah Cool, later told STAT the company would focus at least in part on rare diseases. While “neglected disease” often refers to tropical diseases such as Chagas, leprosy, rabies, dengue, and chikungunya, Anthropic is using the term more broadly to refer to indications that markets and drug companies have neglected, including tropical diseases and rare genetic disorders.” STAT News’ Brittany Trang reports that as part of Anthropic’s rollout of its new science-focused model, the company is also making a play to become directly involved in drug development.
- “Blending together dozens of ingredients, the researchers have synthesized simple cells that feed, grow, reproduce and compete with one another for food. If these cells are not yet fully alive, they have most of the hallmarks of life.” At the New York Times, Carl Zimmer and Marco Hernandez report on the successful creation of simple synthetic cells, a major milestone on the path to creating artificial life.
- “This secret shopper study found that online GLP-1 RA prescription vendors often did not require clinician interaction, relying primarily on patient-reported questionnaires that may not capture important clinical and social history. Several findings suggest limited oversight: multiple GLP-1 RA prescriptions from the same clinicians, prescriptions issued despite missing required photos, and prescriptions issued within 5 minutes or less.” A research article published in JAMA by Chetty and colleagues presents results from a “secret shopper” study that shows prescribing of GLP-1 receptor agonist drugs appears to be receiving lax oversight.
COMMUNICATIONS & Policy
- “…is trust in science really that weak? Researchers studying this have reached some surprising conclusions. From a global perspective, public trust in science and scientists is high, they say. … ‘The idea that there’s a generalized, pervasive lack of trust in science and experts is just completely unfounded in my mind,” says David Bersoff, head of research at the Edelman Trust Institute, a think tank in New York City.’” In a news feature for Nature, Helen Pearson examines recent evidence suggesting that the widely reported crisis of trust in science is a lot more nuanced than some accounts have suggested.
- “…when you see reports of social media engagement spiking about a health myth, don’t overreact. It’s more likely than not that it’s a relatively small group of Americans who already believe it talking to themselves, and they are vastly outnumbered by those who believe the science. What we need to be on guard for are the breakout myths like death panels. And what deserves much more focus is the very large group of Americans in the middle who, just as in politics, are confused by the discussions on the edges and uncertain about what is true and what to do.” An analysis by Drew Altman at Kaiser Family Foundation dissects recent polling data to find that putative opposition to vaccination in the US may be overstated.
- “The primary claim was unchanged in 39.9% of abstracts, minorly revised in 50.0%, and substantially revised in only 10.2%. Hedging shifts were uncommon and asymmetric, with twice as many claims becoming more cautious as more confident (8.4% vs 4.2%). Major revisions were more frequent after long peer review…biomedical papers that were never posted as preprints were retracted at roughly twice the rate of those that were. Together, these data show that the move from preprint to peer-reviewed publication leaves the central claims of most biomedical abstracts intact, indicating that preprints are a reliable source of biomedical research.” An LLM-assisted analysis of biomedical publications by Yin and Rust, available as a preprint from bioRxiv, provides a comparison of published articles that did vs did not receive preprint publication before the final print or epub version.
- “That invisible, geographically diverse workforce digitizing paper records for computers has shrunk significantly—but the actual labor of data entry hasn’t. Big companies just found ways to convert data-entry work from a low-wage job they outsourced to Barbados and Ireland, to something all of us do, for free, every time we enter the doctor’s office or file an insurance claim.” An article published in The Baffler by Michael Waters traces the long history of promises of automation that didn’t so much reduce the need for human labor as cause it to be shuffled somewhere else, out of sight.
