Deskilling Is Not One Endpoint: Five Quantities Behind Assisted Gains and Unassisted Losses in Clinical AI, Why Two Studies of the Same Design Disagree, and Why a Within-Clinician Trial's Control Arm Has Already Been Exposed

Abstract

Clinical AI tools are approved and bought on one endpoint, the performance of a clinician working with the tool, while a second literature asks what happens to the same clinician working without it. The two are usually set against each other as gain against loss. This paper argues that the comparison is malformed, because "deskilling" names at least five different quantities, the studies most often cited measure different ones, and the trials that establish the gain are themselves exposed to the loss. The five are the assisted effect against an unexposed clinician, unassisted performance while the aid is in routine use on other cases, unassisted performance after the aid has been withdrawn for a period, performance when the aid is present but wrong, and the skill a learner acquires when trained with the aid. The colonoscopy evidence illustrates the problem. A multicentre observational study reported a fall in adenoma detection on non-AI procedures from 28.4% to 22.4% after AI was introduced; a prospective three-phase trial found no significant change after removal; and a single-centre study with AI in half the rooms found non-AI detection maintained. The first and third measure the same quantity and disagree, so the conflict is not only a confusion of estimands. The paper then generalises a point the authors of the Polish study made about their own data, where AI-assisted procedures in the post-AI period reached 25.3%, above the concurrent non-AI 22.4% but below the pre-AI 28.4%: when cases are randomised within clinician, the control arm is unassisted work by exposed clinicians, so any spillover from exposure, in either direction, biases the trial estimate of the assisted gain. Facility-level cluster randomisation avoids this. Neither literature reports the share of cases handled without the aid, which is the term that converts any deskilling estimate into a deployment cost. The paper consolidates the published values, states a decision procedure that makes each missing term explicit, and specifies three designs that would settle the open parts.

About the author

Pranay Mahendrakar is an ai specialist and large language model engineer based in Bengaluru, India. He builds production artificial intelligence systems and publishes open-access research on how those systems fail. See all 81 papers by Pranay Mahendrakar, or his ORCID record.