Using ML decision aids can erode workers' decision-making skills: randomized experiments find sizable performance drops when the system is removed, particularly among users who put greater trust in the tool.

Main Finding

Reliance on ML decision aids can causally undermine the development of human decision-making skills. In an incentivized experiment, participants who used ML predictions learned less and showed significant performance drops when the aid was removed; the greater the trust/reliance on the ML outputs, the larger the later skill deficit.

Key Points

• ML decision aids improve short-term decision performance but can produce "deskilling": users stop actively engaging with task-relevant information and fail to form robust procedural knowledge.
• The paper frames learning via the ACT‑R cognitive model: skills develop by converting declarative knowledge (observations, feedback) into procedural production rules through practice. Excessive reliance on ML interrupts this conversion.
• Skill deficits become visible when the system is unavailable (intentional discontinuance, malfunction, or rapid concept drift), since users lack the procedural rules to perform unaided.
• Trust in the ML system strongly moderates the effect: higher trust → more reliance → larger decline in unaided performance.
• Concept drift and sudden, hard-to-anticipate environment shifts (e.g., pandemics) increase the value of retained human skills because ML models may rapidly lose validity; even advanced MLOps cannot always prevent sudden failures.
• The study builds on deskilling/automation literature but focuses specifically on ML decision aids (as distinct from older expert systems) and the acquisition of entirely new cognitive/problem-solving skills.

Data & Methods

• Primary method: a novel, incentivized laboratory experiment designed to identify causal effects and to mitigate endogeneity concerns.
• Task: participants solved logical puzzles where discovering the underlying rule/strategy (procedural skill) increased earnings. Continuous feedback enabled trial-and-error learning.
• Treatment vs control: some participants received ML-generated predictions (decision aid) while others did not. The aid’s error rates were observable to participants (allowing analysis of trust formation).
• Identification strategy: causal inference comes from the experimental assignment and the deliberate removal of the ML aid (system discontinuance) to observe downstream effects on unaided performance.
• Measured outcomes: reliance/trust in the ML predictions, learning/skill acquisition during training, and performance after the ML aid was removed (performance drop as indicator of deskilling).
• Note: exact sample size, demographic composition, ML model type, and statistical estimates are not included in the provided excerpt.

Implications for AI Economics

• Human capital depreciation under automation: Firms adopting ML aids face a trade-off between immediate productivity gains and longer‑term erosion of employee skills. Economic models of automation should account for endogenous human-skill depreciation and its effect on organizational resilience.
• Resilience and risk valuation: The value of human skills increases with the risk of system failure or sudden concept drift. Firms operating in high‑drift environments (fast-changing markets, crisis-prone sectors) should discount returns from automation more heavily or invest more in maintaining skills.
• Designing incentives and institutions: To mitigate deskilling, organizations should (a) design ML aids that encourage active human engagement (e.g., require justification, promote counterfactual thinking), (b) rotate tasks or periodically remove aids to force practice, (c) maintain training programs and human-in-the-loop interventions, and (d) monitor trust formation and reliance to detect unhealthy dependence.
• Product design and regulation: ML system designers and regulators should consider requirements that preserve skill retention (transparency, adjustable automation levels, auditability). Policies encouraging human oversight and retraining subsidies can reduce societal costs of deskilling.
• Labor markets and welfare: Short-term employment productivity gains may mask longer-term reductions in worker versatility and employability. Policymakers should factor deskilling externalities into assessments of AI adoption and labor-market interventions.
• Cost–benefit and investment decisions: Firms must internalize the expected cost of potential discontinuance (retraining, temporary productivity loss). Dynamic investment decisions (how much to automate vs. keep manual) should incorporate the probability and cost of skill erosion and future model failures.
• Research agenda for AI economics: Empirical and theoretical models should quantify skill depreciation rates under different automation regimes, characterize optimal hybrid policies (when to automate vs. preserve practice), and estimate macroeconomic effects of widespread deskilling (e.g., on adaptability, innovation).

If you want, I can (a) draft a short formal model of the trade‑off between immediate gains from ML aids and long-run skill depreciation, or (b) extract specific managerial recommendations for ML deployment (e.g., UI features, training schedules) based on the paper.