The disclosure question has been treated as settled in most institutional policies: tell patients when a model materially influenced their care. New survey data suggests the public agrees with the principle and diverges sharply on the mechanics.

Eighty-six percent of respondents said they would want to be told if an algorithm contributed to a diagnosis. Support dropped to fifty-one percent when the disclosure was described as appearing on every visit summary regardless of the model's role, with many respondents describing that as noise rather than transparency.

The strongest consensus was around recourse. Seventy-nine percent said disclosure without a way to ask for human re-review would be "worse than not being told," a phrasing several respondents volunteered independently in the free-text portion.

Health systems designing disclosure language now face a genuine tension. A blanket notice is easy to implement and legally clean. A graduated notice tied to the model's actual influence is what patients say they want, and requires knowing, per encounter, how much the model mattered — which most deployments cannot currently answer.