AI helps — but not always the person you expected

Marzyeh Ghassemi's group at MIT reports two large randomised experiments in Nature Medicine vol. 32 issue 8 (published online Aug 4, 2026, open access), using a 4×2 factorial design: four explanation formats (bare prediction, GradCAM heatmap, similar-case retrieval, LLM narrative) crossed with two decision orders (human-first vs AI-first). Participants were 623 lay people on binary melanoma classification and 153 primary care physicians on complex differential diagnosis, with 320 medical students as a comparison group.

The findings read like a purpose-built warning about the current trend of patients pointing AI at their own skin photos. For lay users: AI lifted accuracy from 69.7% to 75.8%, but LLM narratives cut both ways — +13.4% when the AI was correct, −21.1% when it was not, amplifying automation bias. For physicians: top-1 accuracy rose 21.5% and the benefit persisted even under incorrect suggestions; LLM explanations helped least on accuracy (+17.7%) yet were the only format that improved confidence calibration. Both groups shared good news — diagnostic disparities across skin tones narrowed (46.9% relative reduction for lay users, 35.6% for physicians) — and bad news: showing the AI answer first significantly increased deference and anchoring. The ordering of a user interface, in other words, matters on the same order of magnitude as the model's accuracy.

Corresponding author Marzyeh Ghassemi (MIT), co-first authors Xuhai "Orson" Xu and Haoyu Hu, with more than 20 co-authors (full text in Nature Medicine). No commercial product is under test — the study evaluates four generic explainability techniques.

A multicentre randomised trial led by Xiaodong Sun's group at Shanghai General Hospital (Shanghai Jiao Tong University), spanning China, South Korea and Poland, published in Nature Medicine on July 24, 2026. The development cohort comprised 1,843 genetically confirmed patients (3,376 eyes); the randomised cohort was 300 people with suspected inherited retinal disease (IRD), allocated 1:1, with 295 analysed (median age 33, 38.6% female). Retina4IRD integrates multimodal retinal imaging and reached 90.4% top-5 accuracy on internal validation and 85.6% externally.

IRD is the archetypal "rare but increasingly treatable" disease family — gene therapies and trial slots keep expanding, yet naming the causative gene demands scarce subspecialty experience. What matters here is not only accuracy: clinicians in the AI-assisted arm reached 88.5% genetic-diagnosis accuracy versus 67.3% in the control arm — a 21.2 percentage-point gap, p < 0.001 — with top-1 accuracy of 37.8% vs 22.4%. More importantly, composite downstream management scores also improved significantly (37.7 vs 28.5, p < 0.001). Most medical-AI trials stop at read accuracy; this one measures what the clinician then did, which is methodologically unusual.

The team spans Shanghai Jiao Tong University, Nanjing University of Aeronautics and Astronautics, Tsinghua University, Yonsei University and several Polish institutions (full text). Retina4IRD is an academic system; no public commercialisation or regulatory filing was found, and none is assumed here.

CMS granted a New Technology Add-on Payment to Aidoc's CARE (Clinical AI Reasoning Engine) Body CT Multi-Triage, reported by AuntMinnie on Aug 13, effective Oct 1, 2026 for three years, covering qualifying Medicare fee-for-service inpatient cases. The device flags suspected urgent findings on chest, abdomen and pelvis CT (with and without contrast) for workflow triage; it received FDA Breakthrough Device Designation in September 2025 and FDA clearance in January 2026 (Aidoc release, Radiology Business, Diagnostic Imaging).

AuntMinnie frames it plainly as "a novel reimbursement decision for AI software built on a healthcare foundation model." NTAP has historically gone to narrow, single-indication algorithms; awarding it to a product that reads multiple urgent conditions off one scan means CMS has effectively recognised multi-task AI as a payable category. The practical clinical consequence: imaging AI shifts from a hospital-funded efficiency bet to a line item with cash flow behind it, sharply lowering the financial barrier to adoption. It also runs straight into the concern researchers raised in STAT's same-day piece: a per-use top-up rewards volume, and when it expires in two to three years the product must justify itself on clinical benefit alone.

Aidoc (Tel Aviv, Israel) is among the market leaders in imaging triage AI and in January 2026 expanded its CT triage platform to 14 indications, 11 of them new. CARE is its foundation-model line. CMS has not published the per-case add-on amount and none is inferred here (original release).

STAT reported on Aug 11 (by Katie Palmer) that US nursing's pushback on clinical AI is becoming institutional. Two flashpoints are named: laid-off nurses at Montefiore in the Bronx questioning administrative AI displacing staff, and nurses at Kaiser Permanente in California striking over AI surveillance of their work. National Nurses United, representing more than 200,000 nurses, is the main actor.

Clinical AI increasingly fails not because the model is inaccurate but because nobody uses it as designed after deployment. Nurses are the frontline users of bedside alerts, deterioration detection and documentation AI — and among the least studied. Story 7 below makes the point concretely: in the UVA deterioration trial, clinicians moved 11% of patients between display and non-display beds, letting clinical judgement override the randomisation outright. STAT's argument is to convert an adversarial relationship into co-design — which is why "who was involved in deployment" is becoming as consequential a variable as model performance.

Ole A. Andreassen of the University of Oslo and 21 co-authors set out the technical foundations, data assets and deployment roadmap of the Nordic AI-Health Initiative in a Nature Medicine Comment (Aug 14, 2026). Participating institutions span Norway, Denmark, Iceland, Sweden, Finland and Estonia, including Karolinska Institutet, the University of Helsinki and Amgen's deCODE genetics. The stated aim is to "enable secure, regulation-compliant access and to deliver generalizable models for responsible AI-driven discovery in medicine."

Nordic longitudinal population registries — cradle-to-grave, multi-generational, linkable — are a globally scarce asset that GDPR and cross-border governance have made hard to use jointly. The significance is not any single model but that this is among the first regional attempts, post-EU AI Act, to build compliance into the data infrastructure itself rather than bolt it on afterwards. The reference value for Taiwan is high: the NHI database has the same longitudinal advantage and the same cross-institutional governance bottleneck (see the Taiwan Angle below). Note this is a Comment, not primary research — no cohort sizes or dated milestones are published.

RSNA announced the 2026 Knee Abnormality Detection AI Challenge on Aug 5, 2026. The dataset holds more than 5,000 knee MRI exams from 16 sites worldwide, with radiology reports in 12 languages. Total prize money is $77,000, including awards for the most efficient models; the challenge runs on Kaggle through Oct 22, winners announced in November and recognised at RSNA 2026 (Nov 29–Dec 3, Chicago). Co-leads are Po-Hao Chen, MD, MBA and Naveen Subhas, MD, MPH of Cleveland Clinic (AuntMinnie).

RSNA calls it "the most real-world challenge yet" precisely because it withholds clean structured labels and forces entrants to learn from messy, variably formatted, multilingual reports. That matters disproportionately for non-anglophone health systems, Taiwan included: most imaging-AI training corpora and benchmarks are English-report-based, and language mismatch is the most commonly underestimated source of performance drop at deployment. The second signal is the efficiency prize — academic benchmarks are finally treating inference cost as a first-class metric rather than optimising AUC alone.

Jessica Keim-Malpass, Jamieson M. Bourque and colleagues at the University of Virginia School of Medicine published a cluster-randomised controlled trial in Scientific Reports (Feb 5, 2026) covering 10,422 inpatient visits on an 85-bed cardiology and cardiac surgery ward. The intervention was CoMET (Continuous Monitoring of Event Trajectories) from Nihon Kohden Digital Health Solutions, which passively displays a risk trajectory refreshed every 15 minutes from machine-learning models over cardiorespiratory monitoring and vital signs. The primary outcome was hours free of clinical deterioration (death, emergent ICU transfer, emergent intubation, cardiac arrest, emergent surgery) within 21 days of admission.

"There was no change in the primary outcome between groups" — passively displaying a risk score did not change patient outcomes. Two methodological details deserve more attention than the headline. Only 5.3% of patients had a deterioration event, so power was structurally thin; and clinicians moved 11% of patients between display and non-display beds, preferentially shifting sicker patients into monitored beds, compromising the randomised design outright. This is exactly the bind flagged in Nature's July 28 comment, "Medical AI has a measurement problem" (Arjun K. Manrai, Harvard Medical School): evidence-based medicine assumes the yardstick is reliable, and in the AI setting that assumption is often shaky. Clinical AI that only displays and does not redesign the workflow and escalation path tends to produce precisely this result.