Screening programs that added AI assistance have generally reported an early increase in recall rates, as radiologists calibrate to a new second reader that flags more than a human colleague would. Whether that increase persists has been the open question.

A program reporting two years of continuous data now offers an answer for at least one setting. Recall rates rose eleven percent in the first two quarters, plateaued, and then declined below the pre-AI baseline starting in the sixth quarter. Cancer detection rate rose modestly and held.

The program's radiology lead attributes the reversal to a feedback loop rather than to the model: readers received quarterly reports on their own agreement with the AI, including cases where they recalled on an AI flag that proved benign. "The model did not get better," she said. "We got better at knowing when to disagree with it."

That interpretation is difficult to verify from the published data, and the authors say so. It is nonetheless the most concrete account yet of the calibration period that most screening programs have been navigating without a map.