pith. sign in

arxiv: 2604.03501 · v3 · pith:AYL2EDHUnew · submitted 2026-04-03 · 💻 cs.HC · cs.AI

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

Michael Caosun , Sinan Aral This is my paper

Pith reviewed 2026-05-22 10:01 UTC · model grok-4.3

classification 💻 cs.HC cs.AI
keywords AI productivityskill erosioncognitive offloadingaugmentation trapdynamic modelworker expertiseproductivity decomposition
0
0 comments X

The pith

Even rational decision-makers adopt AI tools that erode skills, leaving workers less productive in the long run.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper builds a dynamic model where managers decide how intensively to use AI with workers over time. It separates the productivity boost into parts that do not depend on the worker's skill and parts that do. This setup shows that choosing AI for quick gains can rationally lead to a situation where the worker's overall productivity drops below the level before any AI was introduced. When managers focus on short-term results or when skills have value outside the firm, this becomes a trap that harms the worker more than never adopting AI. The same framework predicts that workers with different starting experience can end up on paths where some keep growing skills and others lose them entirely.

Core claim

In a dynamic model of AI usage intensity, the decomposition of productivity effects into an expertise-independent channel and an expertise-scaling channel shows that optimal adoption produces steady-state loss even when erosion is fully anticipated, turns into an augmentation trap under short-termist incentives or external skill value, and allows permanent skill divergence when the tool relies less on existing expertise.

What carries the argument

The dynamic choice model of AI usage intensity over time with a two-channel productivity decomposition separating expertise-independent and expertise-dependent effects.

If this is right

  • Even full anticipation of skill erosion leads to rational AI adoption and steady-state productivity loss for the worker.
  • Short-termist managers or external value of skill convert the loss into an augmentation trap leaving the worker worse off than no adoption.
  • Lower dependence of AI on worker expertise causes experienced workers to reach full potential while less experienced ones deskill to zero.
  • The five regimes from the productivity decomposition separate beneficial from harmful AI deployments and highlight trap vulnerability.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Long-term contracts or skill-maintenance requirements could counteract the trap by aligning incentives with sustained expertise.
  • Industries with heavy AI use may see widening gaps between high- and low-experience workers over time.
  • Empirical tests could track specific skill metrics like problem-solving speed without AI assistance before and after rollout.

Load-bearing premise

Sustained AI use erodes worker expertise over time in a manner that can be separated from the immediate productivity gains and that this erosion dominates the long-run costs.

What would settle it

Track a cohort of workers' productivity on tasks without AI access over several years after introducing sustained AI assistance in their main work; if long-run productivity without the tool falls below the pre-AI baseline, the steady-state loss claim holds.

Figures

Figures reproduced from arXiv: 2604.03501 by Michael Caosun, Sinan Aral.

Figure 1
Figure 1. Figure 1: Total discounted value vs. steady-state value (β = 1, γ = 1, κ = 0.3, δ = 0.1, S¯ = 1). The green curve plots the change in total discounted value V (S¯) − S/δ ¯ , which is positive throughout the adoption region, meaning adoption is always privately rational. The blue curve plots the change in steady-state value V (Sˆ) − S/δ ¯ , which dips below zero in the steady-state loss region (α0 < α ≤ α1) [PITH_FU… view at source ↗
Figure 2
Figure 2. Figure 2: The discount-rate gap (α = 0.5, β = 1.3, γ = 1, κ = 0.3, S¯ = 1). The change in total discounted value (green) is positive for all discount rates above the adoption threshold, meaning adoption is always privately rational. The change in steady-state value (blue) is negative throughout, meaning the long-run position is worse than no AI. The more impatient the decision￾maker is, the worse the loss becomes. C… view at source ↗
Figure 3
Figure 3. Figure 3: Optimal AI usage as a function of skill for the three complementarity regimes. When β > 1, higher-skill workers use AI more; when β < 1, lower-skill workers use AI more; when β = 1, usage is flat. Dots mark steady states. Parameters: α = 1.2, γ = 1.0, κ = 0.3, δ = 0.1, S¯ = 1. Curves shown for β = 1.5 (complementary), β = 1.0 (neutral), and β = 0.5 (leveling). In the long run, workers relying on AI the mos… view at source ↗
Figure 4
Figure 4. Figure 4: Five regions of the (α, β) parameter space (γ = 0.15, κ = 0.3, δ = 0.1, S¯ = 1). Solid line C0: adoption onset; dotted line C1: automation onset; dashed line B: long-run break-even; dash-dot line D: α − γ = S¯, above which full automation yields higher output than the no-AI baseline. Steady-state loss (Region II, pink) is the wedge between C0 and B where adoption is rational but the steady state is worse t… view at source ↗
Figure 5
Figure 5. Figure 5: Persistent skill stratification under β < 1 with (1 − β + 2κaS¯)S >¯ 2γ. Parameters: β = 0.3, S¯ = 5, κ = 0.1, γ = 1, α = 2.5, δ = 0.1. Workers with initial skill below the unstable equilibrium Seq ≈ 1.67 adopt AI heavily and converge to Sˆ = 0. Workers above Seq use little or no AI and converge to S¯. The same technology widens the skill distribution permanently. 2. For any worker with initial skill S0 ∈ … view at source ↗
Figure 6
Figure 6. Figure 6: Productivity, skill, and usage around AI adoption for representative parame￾ter configurations (continuous time, γ = 0.15, κ = 0.3, δ = 0.1, S¯ = 1). Each curve corresponds to a representative (α, β) point from one of Regions I–V in [PITH_FULL_IMAGE:figures/full_fig_p016_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: How κ/δ reshapes the region map. The ratio κ/δ measures how fast skill recovers relative to the decision-maker’s planning horizon. Left: κ/δ = 1.0 (slow recovery, short horizon). Center : κ/δ = 3.0 (baseline). Right: κ/δ = 10.0 (fast recovery, long horizon). The steady-state loss region (Region II, salmon) contracts as κ/δ rises, because skill atrophy is less costly when recovery is fast and the decision-m… view at source ↗
read the original abstract

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.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 1 minor

Summary. The paper develops a dynamic model in which a decision-maker chooses AI usage intensity over time for a worker, trading immediate productivity gains against long-run erosion of expertise. Productivity is decomposed into an expertise-independent channel and an expertise-dependent channel that scales with skill. The model yields three results: (1) fully rational adoption can produce steady-state productivity below the pre-adoption level when front-loaded gains outweigh discounted skill costs; (2) short-termist managerial incentives or external skill value convert this into an 'augmentation trap' that leaves the worker strictly worse off; and (3) when AI productivity depends less on expertise, small incentive differences can drive permanent skill divergence, with experienced workers retaining full potential while others deskill to zero. Deployments are classified into five regimes that distinguish beneficial from harmful adoption.

Significance. If the skill-erosion premise and two-channel decomposition hold in practice, the framework supplies a useful analytical tool for distinguishing productive AI augmentation from deskilling traps and for classifying deployments by vulnerability. The explicit separation of channels and the derivation of regime boundaries constitute a clear theoretical contribution. However, the quantitative predictions rest entirely on the chosen functional forms and parameter values for erosion and productivity; without calibration, validation data, or robustness checks against alternative specifications, the results remain illustrative rather than predictive.

major comments (2)
  1. [Model and Results] The steady-state loss result (first main claim) and the augmentation trap (second claim) are generated directly from the model's chosen functional forms for the skill transition equation and the productivity decomposition P = f(independent) + g(expertise) * h(usage). The manuscript should demonstrate that these outcomes survive alternative specifications, such as multiplicative erosion, threshold-based decay, or recovery with non-use, which can produce interior steady states without net loss.
  2. [Abstract and Model Setup] The abstract and model description provide no empirical calibration, validation data, or robustness checks against real productivity or skill trajectories. Because soundness rests on whether the separable erosion premise matches observed settings, the paper should either supply external benchmarks or explicitly bound the domain of applicability.
minor comments (1)
  1. Notation for the two productivity channels and the five regimes would benefit from a summary table or diagram to improve readability.

Simulated Author's Rebuttal

2 responses · 1 unresolved

We thank the referee for the constructive and detailed comments, which help clarify the scope and robustness of our theoretical framework. We have revised the manuscript to address the concerns about alternative specifications and domain of applicability. Point-by-point responses follow.

read point-by-point responses

  1. Referee: [Model and Results] The steady-state loss result (first main claim) and the augmentation trap (second claim) are generated directly from the model's chosen functional forms for the skill transition equation and the productivity decomposition P = f(independent) + g(expertise) * h(usage). The manuscript should demonstrate that these outcomes survive alternative specifications, such as multiplicative erosion, threshold-based decay, or recovery with non-use, which can produce interior steady states without net loss.

    Authors: We agree that robustness to functional forms strengthens the contribution. In the revised version we have added Appendix C, which re-derives the steady-state and trap results under multiplicative erosion (decay proportional to current skill), threshold-based decay (erosion only above a usage intensity), and partial recovery during non-use periods. The core qualitative outcomes—steady-state productivity below the pre-adoption level and the augmentation trap under short-term incentives—persist across these specifications, although the exact parameter thresholds shift. We retain the original additive forms in the main text for analytical transparency while noting the conditions under which interior steady states without net loss can arise. revision: yes

  2. Referee: [Abstract and Model Setup] The abstract and model description provide no empirical calibration, validation data, or robustness checks against real productivity or skill trajectories. Because soundness rests on whether the separable erosion premise matches observed settings, the paper should either supply external benchmarks or explicitly bound the domain of applicability.

    Authors: This is a purely theoretical model paper; we cannot supply new empirical calibration or validation data without conducting a separate empirical study, which lies outside the current scope. We have, however, revised the abstract and Section 2 to explicitly bound the domain: the framework applies to settings in which AI use can produce cognitive offloading and skill erosion through reduced deliberate practice, consistent with existing empirical findings on AI productivity and expertise maintenance (now cited). We also state that quantitative predictions are illustrative and direct readers to the new robustness appendix for sensitivity to functional forms. revision: partial

standing simulated objections not resolved
  • Supplying empirical calibration, validation data, or robustness checks against real productivity or skill trajectories, as the manuscript is a theoretical contribution without new data collection.

Circularity Check

0 steps flagged

Model derives results from explicit assumptions on erosion dynamics without reduction to inputs by construction.

full rationale

The paper presents a dynamic optimization model that decomposes productivity into independent and expertise-dependent channels and specifies a skill transition process. The three main results, including rational adoption producing steady-state loss, follow directly from solving the model under the stated functional forms and objective. No quoted equations show an output being redefined as its own input, no parameters are fitted to data then relabeled as predictions, and no load-bearing claims rest on self-citations. The derivation remains self-contained within the model's internal logic and stated premises, which is the standard structure for theoretical economic models.

Axiom & Free-Parameter Ledger

2 free parameters · 2 axioms · 0 invented entities

The central claims rest on the untested premise that AI usage causes cumulative skill erosion separable into two channels, plus standard assumptions of rational forward-looking choice and additive productivity effects. No free parameters or invented entities are explicitly listed in the abstract, but the model necessarily introduces functional forms for erosion rates and channel weights that function as fitted or chosen parameters.

free parameters (2)
  • erosion rate parameter
    Rate at which sustained AI use reduces worker expertise; required to generate the steady-state loss result but not given numerical values in the abstract.
  • expertise-scaling weight
    Relative strength of the expertise-dependent productivity channel; determines whether workers diverge in skill or converge to the trap.
axioms (2)
  • domain assumption Worker expertise erodes monotonically with cumulative AI usage intensity.
    Invoked to create the long-run cost that trades off against immediate productivity gains.
  • standard math Manager chooses usage intensity to maximize a discounted sum of productivity net of skill costs.
    Standard dynamic optimization setup used to derive the rational adoption and trap results.

pith-pipeline@v0.9.0 · 5742 in / 1582 out tokens · 31961 ms · 2026-05-22T10:01:22.948758+00:00 · methodology

Review history (2 revisions) →

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

2 extracted references · 2 canonical work pages

  1. [1]

    Acemoglu, D., Autor, D., and Johnson, S. (2026a). Building pro-worker artificial intelligence. NBER Working Paper 34854, National Bureau of Economic Research. Acemoglu, D., Kong, J., and Ozdaglar, A. (2026b). Ai, human cognition and knowledge collapse. Working Paper 34910, National Bureau of Economic Research. Acemoglu, D. and Pischke, J.-S. (1998). Why d...

    work page 1998

  2. [2]

    M., Mollick, E

    Dell’Acqua, F., III, E. M., Mollick, E. R., Lifshitz-Assaf, H., Kellogg, K., Rajendran, S., Krayer, L., Candelon, F., and Lakhani, K. R. (2026). Navigating the jagged technological frontier: Field experimental evidence of the effects of AI on knowledge worker productivity and quality. Organization Science. Forthcoming. Ehsan, U., Passi, S., Saha, K., McNu...

    work page arXiv 2026