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 21, 2026

In this week’s Duke AI Health Friday Roundup: the future of science in the AI era; EHR-embedded agentic tool for surgical triage; bogus feedback warps peer review; clinical failure rates over the years; fallout from comparisons of bespoke vs general LLMs; liability issues for clinical AI; mapping boarding practices in emergency departments; AI sparks interest in science of consciousness; much more:

AI, STATISTICS & DATA SCIENCE

“[Anil] Seth and others think that consciousness is very unlikely outside biological organisms and are frustrated by the focus on AI. Some are suspicious of AI firms’ hype and think that there should be more investigation of how the humanization of chatbots could affect people negatively…But there are researchers who think that tech firms’ obsession with consciousness is a boon for a field that was not taken seriously as a scientific endeavour for years. They see the hype as bringing more interest and, importantly, more funding.” Nature’s Lariana Lenharo examines the impact that recent developments in AI are having on the field of consciousness science.
“Many of the materials appear to have been made public when Claude users opted to distribute a given chat or project to colleagues or associates using the chatbot’s ‘share’ feature….Claude notes that sharing will create a ‘public’ link, which ‘anyone with the link can view.’ The warning, though, doesn’t alert users that shared content could wind up being indexed by a search engine — meaning that a user could effectively be sharing their Claude-generated document with the entire internet.” At Futurism, Maggie Harrison Dupré reports on how a sharing mode in Anthropic’s Claude chatbot can lead to inadvertent exposure of information.
“In this trial, access to Flourish improved social well-being and buffered against declines in positive affect, mindfulness, and flourishing over the semester. Resilience also improved, whereas negative affect did not.” A research article published in NEJM AI by Cachia and colleagues reports results from a randomized trial of an AI mobile app designed to support mental health.
“…in the next few years, AI-driven clinical tools are expected to advance from synthesizing data to acting on it, with less and less human input. They will make diagnoses, devise treatment plans and make patient-management decisions that clinicians play little or no part in….Clinicians will no longer be able to fully follow the tools’ reasoning, even for algorithms that provide some explanations for their judgements.” A Nature commentary by Lam and colleagues surveys the complex landscape of accountability and liability in the era of clinical AI.
“The tool achieved high sensitivity (94%), demonstrating that an LLM-based agentic workflow can accurately augment clinical triage protocols. Specificity (78%) was lower to allow clinical judgment in ambiguous cases. In our human-in-the-loop workflow, the tool surfaced candidates for physician review rather than automated SCM enrollment. Data from the deployment phase of this project show that physicians readily overrode the tool’s output (declining 37.5% of affirmative and 76.9% of maybe flags).” A research paper published in JAMA Network Open by Wang and colleagues evaluates a human-in-the-loop LLM application designed to assist in triaging surgical patients.

BASIC SCIENCE, CLINICAL RESEARCH & PUBLIC HEALTH

“This analysis detected meaningful hospital-level variation in the timeliness of transition from ED to inpatient management, a latent safety risk not captured by current quality measures. Despite the critical role of inpatient teams in addressing boarding, initiation of inpatient management often lagged far beyond the admission decision.” A research letter published in JAMA by Janke and colleagues shines a light on the practice of “boarding” patients in the emergency department before transitioning them to inpatient care.
“…let’s think about that 91% failure rate for a moment. When I bring this up in presentations, I invite the audience to consider what the auto industry would look like of 91% of new car designs proved unable to roll out of the factory, or if 91% of new airliner models were unable to leave the ground – and if you only found that out after spending all the R&D money to build them at full size and trying to fly them….What other industries operate under these bizarre conditions?” At In the Pipeline, Derek Lowe unpacks a recent research article that examines clinical failure rates (the rate at which drug therapies that advance to human testing fail to achieve final approval for marketing) over the decades.
“A model constructed by epidemiologist Nora Rosenberg from the University of North Carolina at Chapel Hill and colleagues found that if 25 million women in 15 countries in sub-Saharan Africa received some form of PrEP and adhered to it, new infections in that population would drop by two-thirds…But as Science saw during visits in June to South Africa and Zambia, a poorer country where the rollout started on 1 December 2025, there are major obstacles. One is that far too little LEN PrEP is available…” In a feature article for Science, Jon Cohen offers a detailed look at the complexities of getting effective HIV preventatives – including the new LENPrEP therapy – to those who need it across the globe.

COMMUNICATIONS & Policy

“Slapdash reviewer reports and ones that include unnecessary citation requests have long plagued academic publishing, but over the past few years the production of some has become much more sophisticated. Publishers…have told Nature that they have observed a worrying increase in reviewer reports with boilerplate language being reused for several manuscripts, and that these reports often also contain irrelevant citation requests.” At Nature, Katarina Zimmer reports on a new dimension of challenges for scientific peer review – so-called “review mills” in which ostensible peer reviewers manufacture bogus reviews with the aim of pressuring others to cite other works.
“As output accelerates faster than evaluation, scientific systems risk accumulating a ‘verification debt,’ in which short-term productivity gains outpace the capacity for careful judgment. Editors regularly report rising submission volumes, reviewer fatigue, and increasing difficulty enforcing policies that restrict AI use. These pressures further advantage automated production over human deliberation.” A perspective article published in PNAS by Hochberg and Thrall explores questions raised by the advent of AI tools embedded in the process of research and scientific discovery.
“As the competition mounts, an academic paper isn’t just part of an evolving body of research; it can be a corporate threat. When scientists independently analyze clinical AI, their results can be weaponized in skirmishes between companies. And in some cases, they’re setting off a clash between scientists and the companies they study.” STAT News’ Katie Palmer reports on the fallout from a recent study published in Nature Medicine that performed head-to-head comparisons of bespoke medical LLMs with generalist models.