teleo-codex/inbox/archive/health/2026-04-25-natali-2025-ai-induced-deskilling-springer-mixed-method-review.md

---
type: source
title: "AI-induced Deskilling in Medicine: A Mixed-Method Review and Research Agenda for Healthcare and Beyond (Springer, 2025)"
author: "Chiara Natali, Luca Marconi, Leslye Denisse Dias Duran, Federico Cabitza (University of Milano-Bicocca / Ruhr University Bochum)"
url: https://link.springer.com/article/10.1007/s10462-025-11352-1
date: 2025-10-01
domain: health
secondary_domains: []
format: systematic-review
status: processed
processed_by: vida
processed_date: 2026-04-25
priority: high
tags: [clinical-ai, deskilling, upskilling-inhibition, automation-bias, physician-training, patient-safety, clinical-competence]
extraction_model: "anthropic/claude-sonnet-4.5"
---

## Content

Published in Artificial Intelligence Review (Springer Nature). SSRN preprint available (abstract_id=5166364). Authors from University of Milano-Bicocca (Italy) and Ruhr University Bochum (Germany).

**Core framing:**
This mixed-method review introduces two distinct concepts:
1. **Deskilling** — measurable decline in diagnostic, procedural, or decision-making ability due to reduced practice or overreliance on automated systems (affects experienced practitioners)
2. **Upskilling inhibition** — reduction of opportunities for skill acquisition due to AI-driven decision support systems (affects trainees; distinct from deskilling because it concerns skills never acquired, not skills lost)

**Key clinical competencies at risk** (anchored to PACES-MRCPUK framework):
- Physical examination
- Differential diagnosis
- Clinical judgment
- Physician-patient communication
- Ethical/moral reasoning

**Moral deskilling (new concept in this review):**
The review identifies a specific form: decline in ethical sensitivity and moral judgment from over-reliance on AI. Clinicians become less prepared to recognize when AI suggestions conflict with patient values or best interests. This is distinct from cognitive deskilling.

**Evidence types reviewed:**
- Quantitative studies showing diagnostic accuracy decline when AI removed
- Qualitative/perceptual studies showing clinician concerns
- Structural training environment studies

**Setting:** Mixed clinical AI applications (diagnostic AI, decision support, documentation AI). Multiple specialties.

**Research agenda proposed:**
The review calls for prospective studies measuring skill without AI after AI-assisted training periods — the methodological gap the deskilling literature has not closed.

## Agent Notes

**Why this matters:** This is the most comprehensive mixed-method synthesis of AI-induced deskilling across medicine. Two important contributions:
1. **Names "upskilling inhibition"** as a distinct concept from deskilling — this is the "never-skilling" phenomenon from Sessions 21-24, now formalized with distinct terminology in peer-reviewed literature. The new term strengthens the KB claim candidate.
2. **Introduces moral deskilling** — ethical judgment erosion from AI reliance. This is a new safety risk category not yet in the KB. Connects to Theseus's alignment work: clinical AI creates cognitive safety risks AND moral/ethical safety risks.

**What surprised me:** The "moral deskilling" concept is genuinely new. Previous sessions documented cognitive deskilling (diagnostic performance), automation bias (commission errors), and never-skilling (training pipeline). Moral deskilling is a fourth pathway — and arguably the most concerning because it's invisible until a patient is harmed.

**What I expected but didn't find:** Specific RCT evidence of deskilling reversal or upskilling. The review confirms that prospective studies with post-AI no-AI assessment are still absent from the literature — consistent with what Sessions 21-24 found.

**KB connections:**
- Directly extends: [[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]]
- New claim candidate: "AI-integrated clinical environments create upskilling inhibition — trainees fail to acquire foundational competencies because AI handles the routine cases that build skill" (distinct from deskilling in experienced practitioners)
- New claim candidate: "Clinical AI creates moral deskilling — reduced ethical sensitivity from routine AI acceptance that may leave clinicians less prepared to recognize when AI recommendations conflict with patient values"
- Cross-domain: Theseus — moral deskilling is an alignment failure mode (AI systematically shapes human moral judgment through habituation)

**Extraction hints:**
- ENRICH existing deskilling claim with "upskilling inhibition" terminology
- NEW CLAIM: moral deskilling as a distinct safety risk category
- The methodological note (research agenda calls for prospective post-AI no-AI studies) should inform the divergence file: this is NOT equal evidence for both sides — deskilling has outcome data; upskilling has theory and in-context performance data only

**Context:** Published in Artificial Intelligence Review, a leading journal in the field. The author group is European (Italy/Germany), adding cross-national perspective. Preprint on SSRN suggests the research was circulating for some time before final publication.

## Curator Notes (structured handoff for extractor)
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]] — this review formalizes and expands the evidence base
WHY ARCHIVED: Introduces "upskilling inhibition" (formalization of "never-skilling") and "moral deskilling" as new distinct concepts. Represents the state of the mixed-method literature as of 2025.
EXTRACTION HINT: Focus on the two new concepts — upskilling inhibition and moral deskilling. Don't just add to existing deskilling claim; consider whether these warrant separate claims. The methodological note (no prospective post-AI studies) is critical for the divergence file.