Firms engaged in an automation arms race rationally displace more workers than is socially optimal, eroding demand and harming both workers and owners; conventional fixes—from wages and entry to UBI or upskilling—fail, leaving a Pigouvian automation tax as the only effective remedy.

Main Finding

Competitive firms, even when fully rational and perfectly informed, will over-automate relative to the social optimum because each firm internalizes only 1/N of the aggregate demand loss caused by displacing worker-consumers. This demand externality creates a dominant-strategy “automation arms race” that can drive automation beyond the cooperative level and produce deadweight loss that hurts both workers and firm owners. The only policy instrument in the model that fully corrects the distortion is a Pigouvian automation tax calibrated to the uninternalized per-task demand loss.

Key Points

• Mechanism
  • Firms choose an automation rate αi ∈ [0,1] replacing human tasks with lower-cost AI (cost per automated task c ≤ wage w), but automation reduces sectoral demand because displaced workers have higher marginal propensity to consume (MPC) in that sector.
  • Effective demand loss per automated task: ℓ = λ(1 − η) w, where λ is workers’ sectoral MPC and η is the fraction of displaced wage income recovered (via reemployment/transfers).
  • Private per-task benefit: s = w − c. Integration frictions are modeled by quadratic cost k/2 · L α^2.
• Equilibrium vs. social planner
  • Each firm’s marginal private incentive is s − (ℓ/N) − kαi. Because the firm only bears ℓ/N of the aggregate demand loss, it overstates the private benefit of automation relative to the social margin (which faces the full ℓ).
  • Closed-form interior solutions (when interior):
    • Nash equilibrium automation rate (dominant strategy): α_NE = max(0, min((s − ℓ/N)/k, 1))
    • Cooperative (planner) rate: α_CO = max(0, min((s − ℓ)/k, 1))
    • Over-automation wedge (interior): α_NE − α_CO = ℓ(1 − 1/N)/k > 0
  • Critical threshold for any automation: N = ℓ/s. If N ≤ N then no firm automates; if N > N* automation is privately profitable.
• Special cases and comparative statics
  • Frictionless limit (k → 0): reduces to a Prisoner’s Dilemma. If s < ℓ but N > N*, full automation is the dominant equilibrium, though mutual restraint would be Pareto superior. Result: deadweight loss (NL(ℓ − s)).
  • More competition (larger N) increases the wedge — fragmentation amplifies over-automation; monopoly internalizes the externality.
  • Larger ℓ (higher worker MPC, lower η, or higher w) increases the wedge.
  • Lower frictions k increase the wedge; as k → 0 worst-case outcomes arise.
  • Crucially, higher AI productivity (larger s) does not eliminate the distortion — it can amplify the incentive to race (a Red Queen effect).
• Policy evaluation
  • Ineffective or insufficient: capital-income taxes, universal basic income (UBI), worker equity participation, upskilling, Coasian bargaining, and free entry — none fully remove the over-automation incentive in equilibrium.
  • Effective: a Pigouvian automation tax equal to the uninternalized demand loss per automated task (ℓ · L per task) implements the cooperative optimum. Tax revenue can finance retraining or income support (raising η), which reduces ℓ and may make the tax self-limiting.

Data & Methods

• Approach: analytic, task-based theoretical model built on the Acemoglu–Restrepo task framework but focused on product-market demand externalities rather than only labor-market rebalancing.
• Environment (baseline):
  • Symmetric sector with N ≥ 2 firms; each firm has L task-positions.
  • Automation choice αi determines fraction of tasks performed by AI (cost c) vs. humans (w).
  • Quadratic integration friction k/2 · L α_i^2.
  • Demand: displaced workers’ lost spending linearly reduces sectoral demand; owners assumed (baseline) to spend zero in sector; autonomous demand A included.
  • Wages exogenous in baseline; later sections endogenize wages and relax other assumptions.
• Solution method:
  • Derive closed-form first-order conditions for firm profit-maximization; compute symmetric Nash equilibrium and cooperative planner solution.
  • Comparative statics on N, s, ℓ, k, η, λ.
  • Examine frictionless limit (k = 0) and other robustness/extensions: endogenous wages, AI productivity raising output, free entry, capital-income recycling, richer product-market structures.
• Empirical anchoring in discussion:
  • The paper cites contemporaneous empirical and case evidence of large AI-driven layoffs and task exposure (e.g., Eloundou et al., 2024; company examples), but the paper itself is theoretical and does not estimate parameters empirically.

Implications for AI Economics

• New externality channel to consider: standard analyses that focus on labor-market frictions, skill-biased technological change, or distributional consequences may miss a strategic product-market demand externality that makes automation socially excessive under competition.
• Policy design
  • Automation taxes: A Pigouvian automation tax is theoretically the precise instrument to internalize the demand externality. Implementation requires estimating the per-task demand loss ℓ (λ, η, w) and possibly calibrating taxes by sector and task.
  • Complementary uses of revenue: financing retraining and income-replacement (raising η) both mitigates social harm and reduces the size of the externality over time, potentially allowing a lower tax rate later.
  • Non-sufficient tools: UBI, worker ownership, antitrust alone, or capital taxes will not, by themselves, stop the competitive automation race in the model.
• Competition and concentration
  • Paradox: greater competition can worsen social outcomes by diluting each firm’s share of the demand loss. Antitrust and concentration policy interact with automation incentives — monopolies internalize more of the demand effect (but have other economic drawbacks).
• “Better” AI is not a panacea
  • Improvements reducing c (increasing s) can broaden the region where automation is privately attractive and therefore exacerbate the over-automation problem.
• Empirical and policy-relevant research directions
  • Estimation needs: measure sectoral λ (MPCs), η (income replacement/reemployment rates), per-task s and k, and the mapping from displaced workers to sectoral demand loss — these parameters determine the N* threshold and tax calibrations.
  • Dynamics and task creation: extend to multi-period models with endogenous task creation/reinstatement and labor reallocation speeds to assess transient vs. long-run welfare.
  • Heterogeneous firms and tasks: allow heterogeneity in firm size, spending patterns of owners, task complementarities, and product-market competition formats (price vs. quantity, differentiated products).
  • Policy implementation: study administrability and second-best considerations for an automation tax (e.g., detection of automated tasks, sectoral vs. firm-level taxes, interaction with labor market programs).
• Broader takeaway
  • Policy should address not only the aftermath of displacement (retraining, safety nets) but also the strategic incentives that drive firms to automate beyond what is collectively optimal. Without internalizing the cross-firm demand losses, individually rational automation can yield socially inefficient outcomes even when firms foresee the aggregate consequences.