The Augmentation Trap: AI Productivity and the Cost of Cognitive Offloading

Experimental evidence confirms that AI tools raise worker productivity, but also that sustained use can erode the expertise on which those gains depend. We develop a dynamic model in which a decision-maker chooses AI usage intensity for a worker over time, trading immediate productivity against the erosion of worker skill. We decompose the tool's productivity effect into two channels, one independent of worker expertise and one that scales with it. The model produces three main results. First, even a decision-maker who fully anticipates skill erosion rationally adopts AI when front-loaded productivity gains outweigh long-run skill costs, producing steady-state loss: the worker ends up less productive than before adoption. Second, when managers are short-termist or worker skill has external value, the decision-maker's optimal policy turns steady-state loss into the augmentation trap, leaving the worker worse off than if AI had never been adopted. Third, when AI productivity depends less on worker expertise, workers can permanently diverge in skill: experienced workers realize their full potential while less experienced workers deskill to zero. Small differences in managerial incentives can determine which path a worker takes. The productivity decomposition classifies deployments into five regimes that separate beneficial adoption from harmful adoption and identifies which deployments are vulnerable to the trap.

Summary

Main Finding

The paper builds a tractable dynamic model showing that AI can raise short-run productivity while causing long-run erosion of worker skill through cognitive offloading. Using a two-parameter decomposition of AI productivity (α = skill‑neutral output; β = skill‑complementary output), it identifies conditions where adoption is privately rational but reduces steady‑state productivity (steady‑state loss) and conditions where misaligned incentives (managerial short‑termism or external social value of skill) produce an “augmentation trap” that leaves workers worse off than without AI. When AI is less dependent on user expertise (low β), deployments can permanently bifurcate workers: experienced workers realize gains while novices deskill toward zero.

Key Points

Model intuition
- Productivity with AI = remaining human contribution + AI contribution.
- AI contribution decomposed as α (independent of user skill) + β·S (scales with worker skill) minus diminishing returns γ·u.
- Skill evolves via learning-from-practice and forgetting-from-offloading: more AI usage displaces practice and reduces skill accumulation.
Analytical structure
- Continuous-time dynamic control problem; Bellman equation yields quadratic value function and linear optimal usage policy u*(S) = u0 + u1·S for interior solutions.
- Steady state skill ˆS and usage ˆu are analytically characterized; comparison benchmark is no‑AI discounted value V_noAI = ¯S/δ.
Main regimes (five qualitatively distinct outcomes in (α, β) space)
No adoption (α below adoption threshold): AI not used.
Steady‑state loss (intermediate α): adoption raises short‑run flow payoffs but reduces long‑run steady‑state value relative to no‑AI — adoption is privately optimal yet harmful in steady state.
Long‑run beneficial adoption (large α or high β): adoption improves both short‑run and long‑run outcomes.
Full automation with long‑run loss (α large enough to automate but not large enough to offset lost skill): usage→1, skill→0, long run worse off.
Full automation with long‑run gain (α very large): automation improves long run too.
Role of β (skill complementarity)
- β > 1: AI complements skill. Optimal usage increases with worker skill (u1 > 0). Feedback is stabilizing: skilled workers use more AI but lose skill and then reduce usage; novices build skill and then increase usage. Tends to protect against long‑run harm.
- β = 1: skill‑neutral AI. Key case showing steady‑state loss: a wedge of α values where adoption is privately optimal but reduces steady‑state value.
- β < 1: AI substitutes for skill. Optimal usage decreases with skill (u1 < 0). Feedback is self‑reinforcing; heavy users deskill faster, potentially producing permanent divergence (experienced workers improve, novices decline to zero).
Augmentation trap
- Occurs when the decision‑maker’s objective is misaligned with worker long‑run welfare (e.g., managers with higher δ) or when worker skill has external value not captured by the decision‑maker. Misalignment expands the parameter region where adoption harms workers relative to no‑AI.
Empirical alignment
- Model is motivated by and consistent with experimental and longitudinal findings: short‑run productivity gains with later degraded judgment/intuitive skill (e.g., “intuition rust,” studies in medicine, programming, education).

Data & Methods

Type: Theoretical, analytical model (continuous‑time dynamic optimization).
Core assumptions and primitives:
- Worker skill Sit ∈ [0, ¯S] evolves via dS/dt = κ[¯S(1 − u) − S], capturing learning from practice and forgetting from offloading; κ is common learning/forgetting rate.
- Per‑period productivity p(S, u) = (1 − u)S + (α + βS − γu)u. Parameters: α (skill‑neutral AI productivity), β (skill complementarity), γ (diminishing returns), δ (decision‑maker discount rate), ¯S (worker potential).
- Decision‑maker chooses u_t ∈ [0,1] to maximize discounted output ∫ e^{−δt} p(S_t, u_t) dt.
Solution approach:
- Derive Bellman equation; interior FOC produces linear policy rule in S.
- Prove value function is quadratic and obtain closed‑form expressions for u*(S), steady states, and thresholds (α0, α1, α2, α3 in β=1 case) that delineate the five regimes.
- Analyze comparative statics for β relative to 1, and for δ (short vs long horizons).
Limitations discussed by authors:
- Model holds (α, β) fixed, but in practice β may decline over time as users disengage (automation bias), so the model may understate harm.
- Uses symmetric learning/forgetting rate κ; qualitative results robust to some asymmetry but specifics could change.
- The paper is not an empirical estimation — it synthesizes experimental/field evidence to motivate parameters and generates testable predictions.

Implications for AI Economics

For firms and managers
- Short‑term optimization can rationally lead to skill erosion; managers with high effective discount rates will overuse AI relative to worker long‑run welfare.
- To avoid the augmentation trap, design workflows that increase β (keep humans in the loop, require judgment, make outputs contingent on user expertise), preserve practice opportunities, and monitor long‑run skills.
- Implement complementary policies: task rotation, explicit practice or training time, periodic skill assessments, and limits on full automation for tasks with high skill externalities.
For workers and labor markets
- Adoption can increase short‑run productivity but may reduce long‑term human capital and career mobility if skill erodes; bargaining and compensation should account for this risk.
- Novices are especially vulnerable in low‑β deployments; firms and educators should ensure learning opportunities remain available (mentorship, staged automation).
For policy and regulation
- Externalities (societal value of skill, safety‑critical domains) justify regulation or incentives to preserve human expertise (minimum human‑in‑loop requirements, mandated training budgets, disclosure of automation risks).
- Evaluate AI deployments not only on short‑run productivity but on long‑run capability outcomes; consider sectoral effects (healthcare, aviation, legal, programming) where deskilling has high social costs.
For research and measurement
- Empirical tests: longitudinal measurement of skill trajectories after AI introduction; heterogeneity by initial skill; identification of β via interaction effects between worker expertise and AI lift; estimating manager vs worker discount rates and link to usage.
- Design interventions to raise effective β (interface constraints, transparency, explanation requirements) and measure whether they prevent deskilling.
Distributional and inequality effects
- Low‑β deployments can amplify inequality by creating a bifurcation: experienced workers capture most gains while novices deskill and stagnate. Small differences in managerial incentives can determine which path a worker follows.

Overall takeaway: AI deployments must be evaluated dynamically. Short‑run productivity gains can mask persistent loss of human capability; the α–β decomposition provides a practical toolkit to classify deployments and to design organizational policies (and regulations) that preserve human capital while capturing AI benefits.

Assessment

Paper Typetheoretical Evidence Strengthn/a — The paper is an analytical/theoretical model without empirical estimation or causal identification from data, so it does not provide empirical evidence for causal effects. Methods Rigormedium — The abstract describes a clear dynamic optimization framework that decomposes productivity into two channels and derives multiple comparative-static and equilibrium results, indicating careful formal analysis; however, theoretical results rest on model structure and assumptions (functional forms, single decision-maker/worker setting, parameter choices) and the paper lacks empirical calibration or robustness tests in the abstract, limiting confidence in quantitative or general conclusions. SampleNo empirical sample; an analytic dynamic model of a decision-maker choosing AI usage intensity for a worker over time, with parameters for immediate AI productivity gains, skill-dependent gains, skill erosion dynamics, managerial time preferences, and external value of worker skill. Themeshuman_ai_collab productivity skills_training org_design GeneralizabilityAbstract model abstraction: single worker–decision-maker setting may not capture firms with many workers or team interactions., No empirical calibration: results depend on unspecified functional forms and parameter values., Ignores labor market responses: hiring, exit, retraining, and reallocation are not modeled., Static technology assumption: AI capabilities and complementarities may evolve over time in practice., Institutional and organizational features (contracts, training policies, regulation) are simplified or omitted., Heterogeneity across tasks, industries, and worker skill types is not explicitly modeled.

Claims (9)

Claim	Direction	Confidence	Outcome	Details
Experimental evidence confirms that AI tools raise worker productivity. Organizational Efficiency	positive	high	worker productivity	0.12
Experimental evidence shows that sustained use of AI tools can erode the expertise on which productivity gains depend (deskilling). Skill Obsolescence	negative	high	worker expertise / skill level	0.12
The authors develop a dynamic model in which a decision-maker chooses AI usage intensity for a worker over time, trading immediate productivity against the erosion of worker skill. Task Allocation	other	high	AI usage intensity decision and resulting worker productivity/skill trajectory	0.02
The tool's productivity effect decomposes into two channels: one independent of worker expertise and one that scales with worker expertise. Organizational Efficiency	other	high	components of AI-induced productivity	0.02
Result 1: Even a decision-maker who fully anticipates skill erosion rationally adopts AI when front-loaded productivity gains outweigh long-run skill costs, producing steady-state loss: the worker ends up less productive than before adoption. Organizational Efficiency	negative	high	steady-state worker productivity (relative to pre-adoption)	0.02
Result 2: When managers are short-termist or worker skill has external value, the decision-maker's optimal policy can produce the augmentation trap, leaving the worker worse off than if AI had never been adopted. Worker Satisfaction	negative	high	worker welfare/productivity relative to non-adoption	0.02
Result 3: When AI productivity depends less on worker expertise, workers can permanently diverge in skill: experienced workers realize their full potential while less experienced workers deskill to zero. Skill Acquisition	mixed	high	long-run worker skill distribution (experienced vs less experienced)	0.02
Small differences in managerial incentives can determine which skill path a worker takes (whether they realize full potential or deskill). Hiring	mixed	high	worker skill trajectory contingent on managerial incentives	0.02
The productivity decomposition classifies deployments into five regimes that separate beneficial adoption from harmful adoption and identifies which deployments are vulnerable to the augmentation trap. Adoption Rate	mixed	high	classification of AI deployment regimes (beneficial vs harmful, vulnerability to trap)	0.02

AI can raise short-term output while quietly eroding human skills, and even farsighted managers may rationally accept a lower long-run productivity once AI is adopted; short-termist incentives or external value of skills can trap workers in harmful deskilling paths.