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 deployed as an “amplifier” can produce clear short-term productivity gains while simultaneously producing asymptomatic, long-term erosion of human expertise, judgment, and professional identity. In a year-long longitudinal study of an AI treatment‑planning system in radiation oncology, the authors document this AI-as-Amplifier Paradox (efficiency + hidden erosion), show how those asymptomatic effects harden into chronic harms (skill atrophy, identity commoditization), and propose a worker-centered response—sociotechnical immunity operationalized via a multi-level framework for Dignified Human‑AI Interaction. A lightweight intervention (Social Transparency, a 4W log: who/what/when/why) helped surface social context and mitigate overreliance without disrupting work.

Key Points

• AI-as-Amplifier Paradox: AI simultaneously enhances task performance (speed, accuracy) and erodes underlying human expertise and agency over time. The erosion can be asymptomatic—undetectable by standard performance dashboards—until it becomes chronic.
• Empirical evidence (radiation oncology): RadPlan deployment produced measurable short-term gains (≈15% shorter planning cycles, improved dose-volume metrics) but clinicians reported “intuition rusting,” approving suggestions without understanding them and losing hands‑on skills.
• Asymptomatic harms → chronic harms: Behavioral drift (habitual acceptance of AI outputs) led to manual skill atrophy, diminished epistemic authority, and worries about becoming “bystanders” in one’s profession.
• Intervention (Social Transparency, ST): Adding clinician-facing 4W logs (who did what, when, why) with privacy protections increased visibility into social reasoning behind plans, helping clinicians calibrate reliance and sustain critical engagement.
• Conceptual contributions:
  • AI-as-Amplifier Paradox (diagnostic lens).
  • Sociotechnical Immunity (a response strategy): detect, contain, recover from hidden harms.
  • Dignified Human‑AI Interaction framework: multi-level mechanisms (early-warning signals, containment actions, recovery routines) that balance institutional quality objectives with worker power and dignity.
• Cross-domain transfer: Framework and interventions evaluated in healthcare and AI-assisted software engineering, indicating broader applicability beyond the case study domain.
• Worker-centered stance: Shifts discourse from “future of work” (productivity-centric) to “future of workers” (expertise, agency, dignity as first-class outcomes).

Data & Methods

• Domain and motivation: Radiation oncology (high-stakes, irreversible decisions; multi-stakeholder workflows). Chosen because small workforce and large per-practitioner impact amplify consequences of skill erosion.
• Case system: RadPlan (pseudonym), a commercial AI-assisted radiation treatment planning tool deployed at a multi-site hospital system (Hospital Alpha).
• Study design: Year-long, longitudinal, mixed qualitative methods capturing native, repeated use of an in-situ AI deployment (rare in the literature).
• Participants: 42 professionals total — 15 radiation oncologists, 12 medical physicists, 7 dosimetrists, 8 hospital administrators. Participants spanned experience levels and had moderate–high AI familiarity.
• Data collection:
  • Usage logs of RadPlan.
  • 52 in-situ think‑aloud sessions (captured hesitation, reliance calibration).
  • 24 semi-structured interviews conducted longitudinally (8 at months 2–3, 10 at months 5–7, 6 at months 10–11).
  • 5 participatory workshops spanning roles.
• Intervention and evaluation:
  • Implemented Social Transparency (clinician-only 4W logs; optional name masking; logs isolated from performance metrics to avoid surveillance).
  • Evaluated ST qualitatively via user feedback and observed behavior changes across the year.
  • Framework tested conceptually in a second domain (AI-assisted software engineering) to assess transferability.
• Ethics & access: Built community trust over months; used synthetic/reconstructed interface figures to protect patient data; informed consent obtained.

Implications for AI Economics

• Rethink value calculus: Standard ROI and productivity metrics (speed, accuracy, throughput) understate long-term costs. AI adoption can create negative externalities by depreciating tacit human capital—costs that materialize later as reduced ability, increased error risk without AI, higher retraining needs, and potential liability.
• Human capital depreciation as an economic variable:
  • Firms should model not only immediate productivity gains but also an amortized human‑capital depreciation rate due to AI-induced skill erosion.
  • Incorporate expected recovery/training costs, monitoring/containment costs, and probability of chronic harm into cost–benefit analyses of automation investments.
• Dynamic complementarity vs. substitution: The findings complicate static complementarity/substitution models. AI may be a short-term complement (raising measured output) but a long-term substitute for tacit skills. Economic models should allow endogenous changes in worker skill and identity over time.
• Labor supply and compensation effects:
  • If AI erodes professional identity and craft, it may alter workers’ reservation utility, career choices, and willingness to accept certain jobs—shaping labor supply for specialized occupations.
  • Employers may face pressure to raise compensation, offer retraining, or provide protections to retain skilled workers who fear deskilling.
• Firm-level governance and transaction costs:
  • Implementing sociotechnical immunity (transparency features, early-warning systems, recovery routines) imposes direct costs but may avert larger downstream liabilities (medical errors, reputation, litigation).
  • Firms must weigh governance/trust-building investments against surveillance and performance-monitoring schemes that may worsen power asymmetries and reduce worker cooperation.
• Regulatory and policy considerations:
  • Economic policy should consider mandates for worker-facing transparency, rights to training and to opt out of certain AI-mediated workflows, and reporting requirements for measures of human-skill impacts.
  • Anticipate new forms of collective bargaining and worker inquiry as mechanisms to internalize long-term costs.
• Measurement and research agenda for AI economics:
  • Develop metrics to quantify asymptomatic harms: human capital depreciation rates, probability and timing of chronic harm, cost of recovery interventions, and impacts on non‑measured outcomes (professional identity, job satisfaction).
  • Longitudinal datasets linking AI usage patterns to subsequent skill assessments, error rates without AI, retraining costs, and turnover would enable structural estimation.
  • Model firm adoption decisions in a dynamic framework where AI accuracy, monitoring, worker resistance, and human capital evolve endogenously.
  • Study cross-sector heterogeneity: where tacit expertise matters most (healthcare, engineering, law) the long-term economic tradeoffs are largest.
• Practical takeaways for economic actors:
  • Investors, managers, and policymakers should require long-horizon impact assessments that include worker-centered outcomes.
  • Fund and implement low-cost sociotechnical immunity measures (transparency logs, peer review processes, scheduled manual-skill refreshers) to mitigate long-term costs.
  • Consider incentives (subsidies, tax credits) for firms adopting worker-protective AI practices and penalties for practices producing measurable human-capital erosion.

Overall, the paper argues that economic evaluations of AI must broaden from immediate productivity gains to include the hidden, temporal, and socialized costs of eroding human expertise and dignity. Accounting for those costs will change optimal adoption, governance, and policy choices.