Found first: a primary source the press has not covered yet.
When a radiology report is available alongside a chest X-ray, MedGemma-27B changes its answer on just 4.26% of trials after the patient's actual image is swapped for a different one, according to a paper submitted to arXiv on September 14, 2026. Remove the report and that figure rises to 20.94%, a difference of 16.7 percentage points. The model is largely reading the report, not the image.
What the source says
Researchers from the criticaldata group ran a paired image-swap audit on 3,199 chest X-ray cases from the MIMIC-CXR dataset, drawn from 293 patients. Each case was queried under 14 questions (13 finding-specific plus one composite); the patient's actual X-ray was then replaced with an image from a different study, and the model was queried again. MedGemma-27B's answer changed on 4.26% of image-swap trials with the report present versus 20.94% without it, a paired difference of 16.7 percentage points (patient-clustered 95% CI 15.6 to 17.7). The effect replicated across two additional model lineages. Code, prompts, and run records are available at github.com/criticaldata/MODALENS.
Why it matters
Any clinical workflow that supplies a prior radiology report alongside a new scan risks getting answers that track the report text, not the live image. A model with near-zero image sensitivity cannot function as an independent second read or surface findings the prior report missed. The study's labels are derived from reports, which limits conclusions about visual correctness in either condition, but the sensitivity gap is large and consistent. Deployment settings that pair text reports with live imaging should treat this asymmetry as a known property of the model before relying on its visual output.