Generative AI's near-zero marginal costs can hollow out the middle of content producers and trigger volatile transitions that briefly collapse parts of the market; when AI-generated information creates strong congestion, further AI expansion can reduce total welfare, arguing for output-focused congestion management rather than laissez-faire input adoption.

Main Finding

Generative AI (GenAI) constitutes a fundamentally non‑convex supply shock: near‑zero marginal costs for standardized content plus an endogenous “information pollution” externality generate a kinked production frontier that bifurcates markets into exit / AI / human segments. This produces a “middle‑class hollow” in quality, a non‑monotonic (collapse → selective recovery) transition driven by search frictions and slow human re‑skilling, and welfare effects that depend critically on pollution intensity (low pollution → monotonic gains; high pollution → inverted‑U with welfare losses at high AI penetration). Optimal policy should therefore shift from input regulation to output‑side congestion management (e.g., Pigouvian taxes on algorithmic volume).

Key Points

• Non‑convex supply shock: AI gives near‑zero marginal cost up to an AI quality cap q̄A, while human production has convex costs (1/2 γ q^2). The global cost is the lower envelope, producing a kink and discontinuities in incentives.
• Information pollution: consumers’ utility includes a congestion penalty Φ(DA) (modeled as β ln(1 + η DA)), making DA (AI volume) an endogenous source of search/cognitive cost.
• Theoretical propositions:
  • Pollution‑Induced Unraveling: if consumers are sufficiently pollution‑sensitive (β > β*), lowering AI marginal costs can reduce aggregate welfare despite lower prices.
  • Middle‑Class Hollow (Lemma): the profit function becomes non‑convex so mid‑quality (just above AI cap) production disappears — a bimodal “barbell” supply distribution.
  • Market Segmentation (Proposition): unique cutoffs partition creators into Exit, Adopt AI, or Human Production.
  • Dynamic stability: the creator distribution follows a Fokker‑Planck (mean‑field) gradient flow whose stationary distribution is a Gibbs/Quantal Response Equilibrium.
• Transitional dynamics:
  • “Shock Therapy”: adoption is non‑monotonic. The economy may undergo a temporary ecological collapse (sharp income/participation drop) before selective recovery as surviving humans reconfigure skills.
  • Asymmetric skill reconfiguration: humans survive primarily by shifting from technical execution (syntax/structure) to semantic creativity (intent/nuance); competing on execution is unviable.
• Welfare tradeoffs:
  • When pollution sensitivity is low, AI expansion raises total welfare monotonically (lower prices + efficiency).
  • When pollution sensitivity is high, total welfare is non‑monotonic: early AI adoption raises welfare, but beyond a point further AI expansion increases pollution costs that outweigh gains.
• Policy implication: laissez‑faire AI adoption can be inefficient. Optimal governance is to internalize information pollution (e.g., Pigouvian tax on AI output/volume or output‑side congestion management and platform design).

Data & Methods

• Three‑layer analytical approach:
• Static vertical‑differentiation model with endogenous pollution term in consumer utility: U(θ,q,p,DA) = θ q − p − β ln(1 + η DA). Participation cutoff ˆθ = (p + Φ(DA))/q.
• Supply modeled as discrete technology choice: Human (convex cost CH(q)=1/(2γ) q^2, qH(s)=α s) vs AI (linear cost CA(q)=cA q + κ for q ≤ q̄A, infinite beyond). Global cost C(q)=min{CH, CA}.
• Dynamic adjustment via Mean Field Game / Fokker‑Planck equation: ∂μt/∂t = ∇·(σ∇μt − μt ∇π(s,μt)), identified as gradient flow of a Free Energy functional F(μ) → stationary μ* ∝ exp(π/σ).
• Agent‑Based Model (“Goldilocks V4.0”) to study transition paths and bounded rationality:
  • Population: NH = 50 human creators, NA = 3 foundational AI models, NC = 1000 consumers, T = 200 periods.
  • Creators choose actions {Stay, Re‑skill, Adopt AI, Exit} via Q‑learning: Q update Q_{i,t+1}(a) = (1−α)Q_{i,t}(a) + α[r_{i,t} + γ max_{a'} Q_{i,t}(a')], exploration ϵ maps to diffusion σ in continuum model.
  • Exit if liquidity depleted or participation falls below outside option.
• Calibration (Goldilocks baseline highlights):
  • AI marginal cost cA = 0.05 × cH (AI ≈ 5% of human cost).
  • AI quality cap q̄A = 0.65 × q̄H.
  • Fixed prompt/fine‑tune cost κ = 0.10.
  • Pollution sensitivity β = 2.0 (tuned to produce inverted‑U welfare).
  • Asymmetric learning rates: λ_tech ≈ 0.12 (fast), λ_creative ≈ 0.025 (slow); human re‑skilling friction η = 0.008.
• Empirical/simulation outcomes reported (selected):
  • Sharp price compression (weighted average transaction price falls ≈ 72.3% in the reported simulation).
  • Emergence of bimodal supply distribution and a “missing middle” region above q̄A.
  • Non‑monotonic temporal dynamics: an initial collapse in creator incomes and participation followed by partial recovery concentrated in high‑skill creative niches.

Implications for AI Economics

• Market structure and inequality:
  • Expect observable bifurcation: a large low‑price AI segment, a smaller high‑quality human niche, and a missing middle. This has distributional consequences (concentration of surplus to winners).
  • Labor markets: roles emphasizing execution/technical routine will be compressed; premium shifts to semantic creativity, strategy, and coordination skills. Policy and training should prioritize this asymmetric re‑skilling.
• Adoption dynamics and policy timing:
  • Because transitions can be non‑monotonic with a transient “valley of death,” passive policy (wait‑and‑see) risks large short‑term welfare losses and persistent exit of mid‑skill creators. Active transitional support (income bridges, retraining, matching subsidies) can reduce the collapse depth.
• Platform and regulatory design:
  • Input‑side rules (copyright/data governance) are insufficient. Regulators should consider output‑side instruments that internalize information pollution:
    • Pigouvian-style taxes or fees on algorithmic content volume or synthetic impressions.
    • Platform design mandates to limit synthetic noise (e.g., quality thresholds, throttling volumes, differentiated ranking for human vs synthetic content).
    • Transparency and labeling to improve discoverability and reduce search costs.
• Empirical research agenda:
  • Measure information pollution and DA: operationalize metrics for synthetic content share, signal‑to‑noise ratios, and consumer search costs (click depth, time to relevant content).
  • Test predictions: (i) emergence of bimodal quality distributions and missing middle; (ii) inverted‑U welfare patterns across markets with different β; (iii) transient declines in creator participation/income during rapid AI diffusion.
  • Heterogeneity: study platform designs and market institutions that mitigate pollution (better recommender exploration, curation subsidies) and their effect on equilibrium segmentation.
• Welfare accounting:
  • Standard productivity measures (price declines, output volumes) can be misleading if they ignore attention/quality externalities. Welfare analysis should incorporate congestion effects on consumer utility and the extensive margin (market participation, variety).

Concise takeaway: GenAI is not a mere cost‑cutting technology — its non‑convex production frontier plus endogenous information pollution reshapes equilibrium structure, generates fragile transition dynamics and ambiguous welfare effects. Policies that internalize pollution and actively smooth transitions (skill policy, platform rules) are central to realizing net social gains.