AI firms’ pricing choices can either stall or accelerate automation: monopolists may exclude users with many complex tasks and create learning traps that slow labor displacement, whereas intense or asymmetric competition expands access or induces specialization and thereby quickens complex-task learning and displacement.

Main Finding

Strategic pricing and market structure of AI providers shape not only AI adoption but the direction and speed of task-specific learning. When providers set access prices under asymmetric user task mixes, they may optimally exclude users with many complex tasks (who are hard to monetize but valuable for learning), producing either a convex acceleration of labor displacement or a “learning trap.” Competition changes these dynamics: sufficiently intense symmetric competition accelerates complex-task learning (via a task‑composition effect), while weak competition can slow it (via data fragmentation). Asymmetric competition can further speed displacement through endogenous provider specialization on complex-task users.

Key Points

• Model setup

  • Users face a continuum of tasks each period; tasks are easy or complex. Each user has private θ ∈ [0,1], the share of complex tasks.
  • AI and labor provide payoffs; in the baseline low-technological state AI is better at easy tasks (be > bd > r).
  • Providers set linear prices; users self-select. Costs to provider depend on task mix (ce per easy task, cd per complex task, cd ≥ ce).
  • Task-based learning: AI improves at easy/complex tasks proportionally to the volume of those tasks performed by paying users; learning efficiencies are ρe and ρd.
  • Baseline providers are myopic (choose price to maximize current profits).

• Monopoly benchmark (main intuition & results)

  • Providers optimally choose a cutoff θ̄t: users with θ ≤ θ̄t buy access.
  • When AI improves at complex tasks, valuations of high-θ users rise and the provider expands access; when AI improves at easy tasks, the provider may restrict access to extract rents from low-θ users.
  • Dynamic outcomes:
    • If complex-task learning is relatively efficient (ρd/ρe > 9/16 in the analytic result), adoption increases over time and eventually becomes convex (accelerating displacement).
    • If easy-task learning is relatively efficient or user-type distribution is strongly right-skewed (most users have few complex tasks), a learning trap can arise: providers persistently exclude complex-task users and complex-task learning remains slow.
  • Analytical expressions: closed-form cutoff in the uniform case and an evolution equation for θ̄t (equation (7) in the paper) characterize the dynamics.

• Symmetric competition (two identical providers)

  • Two opposing effects relative to monopoly:
    • Data fragmentation: each provider has fewer users → slower learning.
    • Task-composition: competition lowers prices → the residual buyers are relatively more complex-task intensive → faster complex-task learning.
  • Net effect: if competition is sufficiently intense, the task-composition effect dominates and adoption/complex-task learning accelerate; if competition is weak, data fragmentation can dominate and slow displacement.

• Asymmetric competition

  • When providers have different initial abilities (one better at complex tasks), endogenous specialization can occur: the complex-task–better provider attracts complex-task users, getting more learning signals and further improving, which accelerates complex-task learning and labor displacement.

• Mapping to real world & assumptions

  • The model relies on plausible contemporary features: non-universal AI adoption, providers collecting type-specific signals (usage metadata, thumbs-up feedback), and pricing menus arising from imperfect knowledge of users’ task mixes.
  • Key simplifying assumptions: myopic pricing baseline, linear pricing, two task classes (easy/complex), task-specific additive learning rules, no endogenous labor supply, and uniform or parametric type distributions in examples.

Data & Methods

• Purely theoretical and computational analysis:

  • Closed-form derivations for screening equilibrium cutoff and its comparative statics (Lemma 1).
  • Dynamic law of motion for AI capabilities be,t and bd,t driven by user composition and learning efficiencies (equations (3)–(4)).
  • Analytical condition (ρd/ρe > 9/16) guaranteeing monotonicity and eventual convexity of adoption in a uniform-type example (Proposition 1).
  • Parametric numerical examples and simulated paths illustrating convex takeoffs, slow takeoffs, and learning traps under different parameterizations (ρe, ρd, costs, initial be, bd, and type distributions).
  • Comparative statics for symmetric and asymmetric competition highlighting the data-fragmentation vs task-composition trade-off and specialization dynamics.

• No empirical estimation or external datasets are used; the paper situates assumptions using recent empirical/industry citations (adoption rates, providers’ use of feedback) to motivate model structure.

Implications for AI Economics

• Market structure matters for the technological frontier
  • Pricing and access decisions are endogenous drivers of which tasks AI learns to perform better. Provider incentives can systematically bias learning toward tasks that are easier to monetize, leaving complex-task capabilities underdeveloped absent corrective forces.
• Competition policy and industry structure
  • Intense symmetric competition can accelerate automation of complex tasks (and hence labor displacement). Regulators and policymakers should note that both monopoly and weak competition can slow diffusion, but for different reasons (monopoly may ration to extract rents; weak competition fragments data).
  • Asymmetric market structure may produce faster specialization-led automation; dominant complex-task specialists could rapidly displace labor in high-value tasks.
• Welfare and policy levers
  • Private-profit maximizing pricing can generate socially suboptimal exclusion of complex-task users (learning externalities), suggesting roles for interventions: subsidized access for complex-task-rich users, data‑sharing mandates, or support for coordinated data commons to mitigate under-provision of learning on socially valuable tasks.
  • Monitoring the composition of paying users (task mixes) and the distribution of prices across providers could be an early indicator of which tasks the industry will automate next.
• Empirical tests and further research directions
  • Empirical validation: correlate provider pricing/access policies and user task composition with model performance improvements across task categories; use provider telemetry (if available) or firm surveys on task usage and adoption.
  • Model extensions to consider: non-myopic dynamic pricing, richer menu/sampling/pricing strategies, endogenous labor responses and reallocation, stochastic learning/noisy feedback, privacy and regulatory constraints on data use, and welfare analysis with social planner interventions.
• Takeaway for stakeholders
  • Firms, policymakers, and researchers should treat pricing and access design as a core determinant of AI’s trajectory: who gets access matters not only for near-term adoption but for which capabilities the technology develops and how quickly labor is displaced.