Study: endoscopists accustomed to using AI during colonoscopies detected fewer precancerous growths after the tech was taken away, raising “deskilling” concerns
- The adenoma detection rate (ADR) in more than 1,400 non-AI assisted colonoscopies dropped 6 percentage points …
Context & Ripple Effects
Clinical AI coverage has largely emphasized higher detection: AI-supported mammography increased cancer detection in a Swedish study, while radiology has become the dominant destination for cleared medical AI software. This colonoscopy result adds an operational counterpoint: performance with the tool may not translate cleanly to performance without it.
The finding matters because providers are increasingly deploying AI to augment clinical work, including Mayo Clinic’s effort to use AI across radiology workflows. It shifts attention from validation of an algorithm’s assisted results to how clinicians retain baseline diagnostic skill.
First-order effects
- Among endoscopists accustomed to AI assistance, non-AI colonoscopies in the study had an adenoma detection rate about 6 percentage points lower across more than 1,400 procedures.
- Hospitals using colonoscopy AI must treat tool unavailability as a clinical-performance risk, not merely an IT interruption.
Second-order effects
- Clinical leaders and AI vendors face pressure to measure performance both with and without assistance, and to build training or review practices that preserve unaided detection ability.
- Procurement cases based only on assisted detection gains become less complete; buyers must weigh those gains against dependence, downtime procedures, and clinician training needs.
Third-order effects
- If replicated, the result would make human-factors monitoring a core part of clinical AI deployment, alongside algorithm accuracy and workflow efficiency.
- Medical AI governance may increasingly distinguish augmentation that strengthens clinician judgment from systems that create capability dependence, especially in high-volume screening settings.
The trend: Healthcare AI is moving from proving that assistance can improve detection toward governing the long-term interaction between algorithmic support and clinician capability.