AI accelerates clinicians’ work in the short run but slowly dulls diagnostic intuition, causing skill atrophy and identity commoditization among cancer specialists; the authors propose a sociotechnical framework to detect and reverse such erosion while preserving human expertise.

Main Finding

AI acts as an “amplifier” that simultaneously enhances short‑term performance and erodes the tacit skills, judgment, and professional identity that underlie long‑run human expertise. In a year‑long study of AI use by cancer specialists, initial productivity and operational gains concealed a slow, asymptomatic decline in clinicians’ intuition and decision skills (“intuition rust”), which progressed into chronic harms (skill atrophy, identity commoditization). The authors propose a sociotechnical framework for “dignified Human–AI interaction” that embeds dual‑purpose mechanisms to preserve institutional quality while empowering workers to detect, contain, and recover from AI‑induced skill erosion.

Key Points

• AI‑as‑Amplifier Paradox: AI magnifies both performance and underlying vulnerabilities — it can boost throughput while weakening the human capabilities that sustain high‑stakes judgments.
• Intuition rust: Skill degradation can be gradual and asymptomatic; users may not notice loss of tacit judgment until harms manifest.
• Chronic harms identified:
  • Skill atrophy: erosion of domain expertise and pattern recognition built through practice.
  • Identity commoditization: professionals’ roles become standardized and commodified around AI outputs, undermining autonomy and professional dignity.
• Sociotechnical immunity: The proposed remedy is a framework that integrates institutional quality goals with worker empowerment — mechanisms that simultaneously maintain safety/performance and preserve human expertise and agency.
• Cross‑domain relevance: Authors evaluated and illustrated the framework in healthcare (oncology) and software engineering, suggesting the paradox and mitigation strategies generalize across high‑skill domains.

Data & Methods

• Study design: A longitudinal investigation spanning one year focused on AI use in a high‑stakes clinical setting (cancer specialists). The paper traces behavioral and skill changes that emerged over time with routine AI integration.
• Evidence types (as reported):
  • Longitudinal observations of workplace AI usage patterns.
  • Qualitative data documenting clinicians’ decision processes, self‑reports, and evolving interactions with AI tools.
  • Analytic comparison of outcomes, workflows, and professional practices before and after sustained AI adoption.
  • Cross‑domain evaluation applying the proposed framework to healthcare and software engineering contexts to test transferability.
• Analytic approach: The authors synthesize qualitative longitudinal evidence to surface latent, asymptomatic effects (intuition rust), characterize downstream harms, and derive design and institutional interventions; they then map these interventions to dual technical and organizational mechanisms.

Note: The paper emphasizes longitudinal, qualitative insight into tacit skill dynamics rather than short‑term productivity metrics alone.

Implications for AI Economics

• Human capital depreciation risk: Widespread AI adoption may temporarily raise measured productivity while reducing the accumulation and maintenance of tacit human capital, lowering long‑run labor quality and increasing dependence on AI systems.
• Misleading productivity signals: Firms and markets that reward near‑term throughput gains risk underinvesting in practices that sustain worker expertise; measured GDP or firm performance gains could mask growing systemic fragility.
• Labor bargaining and rent allocation: As expertise erodes and roles become commodified, workers’ bargaining power may decline, altering wage dynamics and rent distribution between capital (AI owners) and labor.
• Task allocation and skill complementarities: The paradox complicates models of automation versus augmentation — tasks that appear complementary in the short run can become substitutes over longer horizons as human capacity atrophies.
• Externalities and safety: Tacit‑skill erosion creates negative externalities (quality degradation, increased systemic risk) that markets may not internalize, signaling a role for regulation or industry standards.
• Policy and firm responses:
  • Monitor tacit skill trajectories: Invest in metrics and audits that detect skill erosion (not just output).
  • Dual‑purpose interventions: Design AI deployment rules that simultaneously preserve quality and encourage skill maintenance (e.g., rotation, human oversight requirements, periodic unguided decision tasks, documentation and knowledge retention incentives).
  • Labor protections and co‑governance: Build institutional mechanisms (worker participation, review boards, retraining mandates, compensation for maintaining non‑AI skills) to preserve worker agency and long‑run human capital.
  • Rethink incentives: Align firm incentives with long‑run competence maintenance — e.g., performance metrics that reward human expertise preservation and resilience.
• Research agenda: Develop quantitative measures of tacit skill erosion, model the long‑run macroeconomic impacts of AI‑induced human capital depreciation, and experimentally evaluate sociotechnical interventions across sectors.

Overall, the paper reframes AI’s economic role: beyond productivity amplification, AI reshapes the dynamics of human capital accumulation, agency, and bargaining — requiring coordinated sociotechnical and policy responses to avoid chronic harms that undermine long‑run economic value.