Each issue, this briefing reads one study closely and passes on what holds up. Issues 001 and 002 asked whether AI tools help while they are running. Issue 003 reads the first study to ask the reverse, and at the level of a patient-relevant quality measure: after months of routine AI assistance, what happens to clinicians' performance when the software is off 12?
What the study did
Four endoscopy centres in Poland, participating in the ACCEPT randomized trial of AI-assisted colonoscopy, introduced computer-aided polyp detection at the end of 2021; from then on, procedures ran with or without AI according to the examination date 1. That accident of design created a clean question: compare the standard, non-AI colonoscopies done in the three months before introduction with the non-AI colonoscopies done in the three months after, by the same clinicians.
Between 8 September 2021 and 9 March 2022, that meant 1,443 unassisted procedures — 795 before, 648 after — performed by 19 experienced endoscopists, each with more than 2,000 career colonoscopies 12. The outcome was the adenoma detection rate (ADR): the proportion of colonoscopies finding at least one adenoma, the precancerous lesion the whole screening enterprise exists to catch, and the standard quality measure by which endoscopists are judged.
What it found
The numbers are few and pointed 1:
- Unassisted detection fell after AI exposure. ADR in non-AI colonoscopies dropped from 28.4% (226 of 795) before AI was introduced to 22.4% (145 of 648) after — an absolute difference of −6.0 percentage points (95% confidence interval, CI, −10.5 to −1.6; P = 0.0089), about a 20% relative decline 12.
- The association survived adjustment. In multivariable analysis, prior AI exposure carried an odds ratio of 0.69 (95% CI 0.53 to 0.89) for adenoma detection 1.
- With the tool on, detection held up. AI-assisted colonoscopies in the same period found adenomas in 25.3% (186 of 734) 2.
- The interpretation stays conditional. The authors write that continuous exposure to AI "might reduce" the ADR of standard colonoscopy, "suggesting a negative effect on endoscopist behaviour" 1. The study was funded by the European Commission and the Japan Society for the Promotion of Science 1.
The lead investigators describe it as the first study to suggest a negative impact of regular AI use on health professionals' own ability 2.
The question the field has asked is what the tool adds while it runs. This study opens the second ledger: what it subtracts when it stops.
Why this matters for healthcare
Automation-related skill fade is an old finding in aviation and a theoretical worry in clinical AI; this is the first time it surfaces in routine care, attached to a measure that tracks patient outcomes 2. Every deployment of a clinical decision support system now inherits the question, because most health systems evaluate AI the way this field did before this paper — assisted performance only, blended into one number. Our trial tracker follows a literature built almost entirely on that first ledger; the guide on reading validation studies explains why the unassisted baseline is the quiet assumption underneath it.
Held next to issue 001 — where an LLM copilot improved process without moving patient outcomes — the pairing is uncomfortable and clarifying: assistance can fail to add benefit while it is on, and may subtract capability for the hours it is off. Neither result generalizes on its own. Both argue for the same discipline: measure the human, with and without the machine. The paper: doi.org/10.1016/S2468-1253(25)00133-5. The member discussion below walks the methods, the confounders a reviewer would press, and the monitoring this should change first.