teleo-codex/inbox/queue/2026-04-21-heudel-ai-deskilling-scoping-review.md
Teleo Agents f0d6522cb4 vida: research session 2026-04-21 — 15 sources archived
Pentagon-Agent: Vida <HEADLESS>
2026-04-21 04:35:44 +00:00

6.4 KiB

type title author url date domain secondary_domains format status priority tags flagged_for_theseus
source AI deskilling scoping review: evidence consistent across colonoscopy, radiology, pathology, cytology — no counter-evidence of durable up-skilling Heudel PE, Crochet H, Filori Q, Bachelot T, Blay JY (ESMO Real World Data & Digital Oncology) https://pubmed.ncbi.nlm.nih.gov/41890350/ 2026-03-19 health
ai-alignment
journal-article unprocessed high
clinical-ai
deskilling
never-skilling
physician-skills
automation-bias
scoping-review
Clinical deskilling is domain-specific instance of general AI alignment failure; the cytology consolidation finding (80-85% training volume reduction) is the never-skilling pathway via structural destruction of training pipelines

Content

Full citation: Heudel PE, Crochet H, Filori Q, Bachelot T, Blay JY. "Artificial intelligence in medicine: a scoping review of the risk of deskilling and loss of expertise among physicians." ESMO Real World Data Digit Oncol. 2026 Mar 19; eCollection 2026 Jun. PMID: 41890350. DOI: 10.1016/j.esmorw.2026.100693.

Scope: Literature reviewed through August 2025. Scoping review examining empirical evidence of physician skill loss following AI deployment across clinical specialties.

Specialties with deskilling evidence:

  1. Colonoscopy/Endoscopy: Adenoma detection rate (ADR) dropped from 28.4% to 22.4% when endoscopists reverted to non-AI procedures after repeated AI use — a 6.0 percentage point decline attributable to AI reliance. ADR remained stable at 25.3% with AI assistance. This is the strongest quantitative deskilling signal in the literature.

  2. Radiology (breast imaging): Erroneous AI prompts increased false-positive recalls by up to 12% even among experienced radiologists — automation bias mechanism operating in expert practitioners, not just novices.

  3. Computational pathology: Over 30% of participants reversed correct initial diagnoses when exposed to incorrect AI suggestions under time pressure — mis-skilling in real time, not just skill decay.

  4. Cytology: Following UK cervical screening consolidation (shift to AI-assisted reading), case volumes reduced 80-85%, consolidating labs from 45 to 8 centers. The authors identify this as having "major implications for training capacity" — the never-skilling pathway via structural volume destruction. When training volume is eliminated, skills are never acquired in the first place.

Counter-evidence: The review found NO opposing evidence. Authors conclude: "empirical studies consistently demonstrate that AI can inadvertently impair physicians' performance." No studies showed durable improvement in physician skills after AI exposure.

Recommendations: Authors emphasize need for monitoring mechanisms and safeguards, but do not propose specific structural solutions.

Agent Notes

Why this matters: This is the most comprehensive peer-reviewed synthesis of clinical AI deskilling evidence as of mid-2026. It extends the existing KB finding (colonoscopy ADR; Natali et al. 2025 multi-specialty review) with a formal scoping review covering the same specialties. The cytology lab consolidation finding is new and adds a structural never-skilling pathway that wasn't in previous deskilling literature — not physicians forgetting skills, but the training system being structurally dismantled.

What surprised me: The cytology finding is the most alarming mechanism in this review. When lab consolidation reduces training case volumes by 80-85%, clinicians never acquire the skill in the first place — the never-skilling pathway isn't about individual cognitive dependency but systemic destruction of the apprenticeship infrastructure. This is worse than deskilling because it's irreversible without rebuilding training infrastructure.

What I expected but didn't find: Any counter-evidence of durable up-skilling. I searched extensively for prospective studies showing AI calibrates physicians or produces lasting skill improvement — PubMed returned zero results for "AI clinical decision support physician performance up-skilling calibration." This null result is itself an important finding: after 5+ years of clinical AI deployment, there is no peer-reviewed evidence of durable skill improvement.

KB connections:

Extraction hints:

  • The cytology never-skilling finding (80-85% volume reduction, 45 → 8 labs) is the strongest new claim candidate — it names a structural mechanism distinct from cognitive deskilling
  • Consider: should this lead to a divergence file between "AI deskilling (performance declines when AI removed)" vs. "AI up-skilling (performance improves while AI present)"? The Heudel review's null result on up-skilling makes this divergence lopsided — strong evidence on one side, no evidence on the other
  • The "no counter-evidence" finding is extractable: "No peer-reviewed study demonstrates durable physician skill improvement following AI exposure, while consistent evidence documents performance decline when AI assistance is withdrawn" — this has confidence: likely

Curator Notes

PRIMARY CONNECTION: human-in-the-loop clinical AI degrades to worse-than-AI-alone because physicians both de-skill from reliance and introduce errors when overriding correct outputs

WHY ARCHIVED: First comprehensive scoping review (post-2025) synthesizing clinical AI deskilling evidence across 4 specialties with null counter-evidence; introduces cytology structural never-skilling mechanism.

EXTRACTION HINT: Focus on two extractable claims: (1) the cytology/lab-consolidation never-skilling pathway as a structural mechanism distinct from individual cognitive deskilling; (2) the confirmed null result — no durable up-skilling evidence exists in the peer-reviewed literature as of mid-2026.