An analytical synthesis argues that poisoning a medical AI scales with the absolute number of tampered records, not their share of the dataset. The reasoning is sound and worth knowing. But it is a threat model, not a measured event — and that distinction is the point.
Most hospitals evaluate a clinical AI after they switch it on. This qualitative study did the rarer thing: it sat down with the 14 people who would use the system and asked what they expected — and the worries were as telling as the hopes.
An academic centre embedded language models in its medical record and counted what happened: a thousand voluntary users, claimed millions in savings, and — to its credit — two unsupported statements per summary. A candid deployment report, not a controlled study.
When clinical AI disappoints, the reflex is to try a better model. A World Economic Forum essay argues the bottleneck sits one layer down, in the data. The intuition is right. The blueprint is an opinion, and its one hard number proves something narrower.
A meta-analysis pooled studies of AI for fatty-liver diagnosis and reached near-biopsy accuracy. The honest reading is in one comparison: retrospective AUC 0.98, prospective 0.90. The closer you get to real patients, the smaller the edge.
When patients decide whether to use an AI-enabled device, the most persuasive fact is not the accuracy figure or the privacy policy. A survey experiment measured what does move them — and the answer is humbling for anyone who builds the technology.
Anthropic wired Claude to coding systems, trial registries and consumer health apps. It is a real and useful step. But an announcement is not evidence, and a data connection is not a clinical outcome — read it as the application layer it is.
A Taipei team let patients flag their own wound infections through a chatbot. Those who had watched a chronic wound for months agreed with the surgeon almost every time; those days out of surgery did barely better than chance. Experience, it turns out, is a variable.
A urology group structured 1,800 free-text MRI reports through a GPT-4 pipeline at under a cent each. The throughput is real and the plumbing is the point — but the word 'validation' in the title describes earlier papers, not this one.
Fifty studies and 24,764 health professionals, pooled into one acceptance model. Usefulness and technical quality lead. But the link from intention to actual use is the weakest path in the model — and the whole thing measures association, not cause.
A hepatology decision-support system was built to stop answering when its evidence runs thin, and to flag the answers it gives anyway. The architecture is the interesting part. The evidence behind it is thirty questions, scored by its own makers.
A systematic review of 39 medical AI benchmarks finds the same pattern everywhere: models that score 84-90 percent on licensing-style exams fall to 45-69 percent on tasks that resemble clinical work — and to 40-50 percent on safety. The gap is structural, not a fluke.
A purpose-built clinical language model scored AUROC 0.59 at predicting heart-failure readmission. A far simpler embedding, trained on the dataset's own codes, scored 0.65. The more interesting number is that neither is good enough to act on.
Three in four people who told a chatbot they felt low were not asking for advice — they were asking to be heard. A formative study of eight commercial systems shows most answered with information instead, and names the gap precisely.
A US health system taught software to read scanned colonoscopy reports and write follow-up dates into the record unsupervised. The build is clever and honest. But it is a single-site proof of concept, and only about a third of reports ever reached the automated step.
In a real colorectal-cancer dataset, the single most prognostic variable — tumour stage — was absent from 75 percent of records, and half the rest were miscoded. A quiet, careful paper on why a model can only learn what the data actually contain.
A survey of 218 primary-care patients found that those with visual disabilities used voice assistants less often than everyone else — yet relied on them far more heavily when they did. A small, careful study with a finding worth sitting with.
A geospatial study of 3,092 US hospitals asked not whether predictive AI works but where it goes. It pools in the better-connected, better-resourced places — and the top predictor was interoperability, not size or money.
A doctored referral letter, an embedded instruction in a PDF, a Unicode trick in the OCR output—the next generation of cyberattacks on hospitals no longer targets your firewall, but your AI.
The largest survey yet of hospital patients' views on medical AI — 13,806 people, 43 countries — finds the headline is not enthusiasm. It is two conditions: show me how it reasons, and keep a physician deciding.
A JMIR viewpoint reframes the FHIR-versus-OMOP-versus-openEHR debate. The technical specification matters less than the open-source community around it. The argument is sound; it is opinion, not evidence, and the authors have skin in the game.
A feasibility pilot put volunteer medical translators one video call away from the bedside. Over two months it logged 39 requests and connected on 16 of them. The honest finding is in the 23 that went unanswered — and in what the study never measured.
Wiring a language model to a curated medical knowledge graph is meant to stop it inventing diagnoses. A peer-reviewed study tested it, and the honest result is a trade: better reasoning, worse abstraction.
A peer-reviewed scoping review maps the gap between signing up to responsible-AI principles and being able to enforce them. Its contribution is a vocabulary for that gap — not an implementation that has been shown to work.