Where a reporting channel existed, nurses used it 86.2% of the time. Fear of consequences predicted nothing. This is not a culture problem but rather an absence of infrastructure.
Every postmarket surveillance framework being designed right now, in the United States and elsewhere, rests on an assumption that when a clinician sees something go wrong with an AI tool, there is somewhere to report it or act on it.
A national survey of 1,189 practicing US nurses, run in June and July 2026 by Gregory Carter and colleagues at the Indiana University School of Nursing, tested that assumption.
424 of the 1,189 had encountered an AI-related patient safety concern. Of those 424, 83 had no formal channel through which to report it. 1 in 5.
The gap is not distributed randomly. It tracks distance from the hospital. No reporting channel: 15.1% in hospitals, 23.9% in home health and hospice, 31.6% in outpatient and ambulatory, 43.8% among public health nurses. Nurses working outside hospitals carried 2.5 times the odds of having nowhere to report, OR 2.54, 95% CI 1.54 to 4.19. Settings not designated high-risk carried 2.3 times the odds.
Here is the finding that changes what you do about it. Where a channel existed, nurses used it. The base rate for reporting was 86.2%. Perceived severity was the only independent predictor of the remainder, adjusted OR 2.47. Fear of consequences predicted nothing at all, P = .401.
So this is not a culture problem or intimidation, and it is not a speak-up problem to be solved with training. This is an absence of infrastructure, which means it is buildable.
It also means something less comfortable. If severity is what drives reporting, then the failures that get reported are the dramatic ones. The gradual drift, the slow miscalibration, the model that degrades over 8 months rather than failing on a Tuesday: those are the failures a lifecycle framework exists to catch, and they are the ones least likely to be filed.
One limitation the paper cannot resolve. The response options were binary. Whether the 4 in 5 who do have a channel file into a general incident system, a dedicated AI pathway, or an email to the vendor is unknown. Those are 3 different instruments reaching 3 different places, and only one of them reaches a regulator.
An institution that deploys clinical AI without a reporting route has not built a safety program. It has built a surveillance framework with no afferent limb.
Before the model, build the communication channel.
Source. Carter G, et al. National cross-sectional survey of AI-related patient safety reporting among US nurses. Journal of Nursing Care Quality, 10 September 2026. doi:10.1097/NCQ.0000000000001011