Generative AI has erased the scarcity of polished expert deliverables, collapsing competence signals and driving price pooling and exit among high-quality providers; outcome-contingent liability—verifiable warranties with sufficient damages—can re-establish differentiation, but only above a minimum ticket size and subject to insurance and enforcement constraints.

Main Finding

Generative AI compresses the observable quality gap between cheap machine-produced artifacts and what buyers can perceive, collapsing production-side competence signals in expert-service markets. When that discernible headroom falls below a critical threshold, artifact-based separation disappears: prices pool, high-competence providers lose their premium and some exit. An outcome-contingent signal — a warranty/liability promise that is verifiable ex post — is invariant to AI capability and can fully restore separation whenever the expected collectible damages cover the buyer’s value at risk (formal condition: φ·D ≥ v). Provenance certification (type-independent cost) cannot substitute. Liability signaling, however, has institutional requirements (minimum ticket size given fixed litigation costs; the seller’s retained risk under insurance matters).

Key Points

• Classification of signals:
  • Production-side signals: costly artifacts presented pre-transaction (reports, polished proposals). Cost is collapsed by generative AI.
  • Outcome-contingent signals: warranties/damages paid upon failure. Cost depends on realized outcomes and seller competence and is unaffected by AI artifact costs.
• Formal collapse result:
  • Define discernible headroom η = s̄ − a (buyer’s maximum distinguishable quality minus AI frontier).
  • Separation via artifacts exists iff η ≥ η* = (Δq · v) / k_L (Δq = q_H − q_L; k_L marginal cost parameter).
  • If η < η*, no separating equilibrium in production-side signals exists → pooling prices and adverse selection (Akerlof lemons dynamic).
• Liability restores signaling:
  • A warranty paying damages D with ex-post verification probability φ restores full separation for any AI level if φ·D ≥ v (the buyer’s value from a solved problem).
  • Intuition: expected liability cost depends on actual competence (q_θ) and outcome verification, not on artifact-production cost.
• Provenance certification (e.g., attestations of human origin, C2PA) is type-independent and therefore cannot restore separation; it certifies process, not competence.
• Institutional extensions:
  • Minimum ticket size: fixed blocks of civil-procedure costs create v_min below which liability promises are not credible/enforceable, so small engagements will pool.
  • Insurance: the effective signal is the risk the seller retains (or is individually priced on). Under pooled premiums, a promised D may not be signal-effective; experience rating gradually restores signal value on insurers’ information timescale.
• Simulation evidence (agent-based Monte Carlo):
  • Baseline calibration: AI capability shock erases ~20 percentage points of competence premium and reduces high-type participation (e.g., 88% → 68%) when only artifacts are available.
  • When liability contracts are available, both premium and participation are sustained; mimicry by low types becomes self-extinguishing because payments are tied to realized outcomes.
• Empirical agenda and falsification:
  • The paper proposes a preregistered, choice-based conjoint experiment (200–300 B2B decision-makers in the German-speaking expert-services market) varying liability × AI disclosure × service attributes to measure monetary willingness-to-pay and test the theory.
  • Falsification condition: if explicit warranties/liability assumption generate no price premium over several years while prices of pure information services change as predicted, the model is wrong.

Data & Methods

• Analytical model:
  • Market for a credence good with two seller types θ ∈ {H, L}, solve probabilities q_H > q_L.
  • Buyers value a solved problem at v; sellers choose participation, artifact quality s ∈ [0, s̄], and whether to offer a warranty w (pay damages D upon verified failure).
  • Artifact production cost: c(s, θ; a) = k_θ · max{0, s − a}; a is AI frontier (artifact quality cheaply attainable); headroom η = s̄ − a.
  • Warranty cost to type θ: κ_θ(D) = (1 − q_θ) · φ · D (φ = probability failure is verifiably detected). κ_θ is independent of a.
  • Timing: types choose actions → buyers observe s and w and set competitive price given posteriors → outcomes realize and verified failures trigger damages.
  • Solution concept: Perfect Bayesian Equilibrium; analytical propositions (1–7) proven in Appendix A.
• Simulation:
  • Agent-based Monte-Carlo implementation with finite populations, adaptive participation dynamics, and low-type experimentation.
  • Calibration is stylized to illustrate dynamics; shows quantitative effects described above.
  • Code and numerical outputs openly available: https://doi.org/10.6084/m9.figshare.33106946 (direct figshare link provided in preprint).
• Empirical proposal:
  • Pre-registered choice-based conjoint varying (liability offered / not), (AI disclosure / human) and other service attributes, eliciting monetary WTP from B2B buyers in German-speaking markets; preregistered falsification criteria described.

Implications for AI Economics

• Mechanism clarity: the paper formalizes a simple comparative static — technology that drives the cost of production-side signals toward zero destroys those signals’ informational content. This reframes observed price declines and trust penalties as two sides of one signaling collapse.
• Market structure and pricing:
  • Expect pooling and premium compression in expert-service segments where buyer value rests on artifact signals that become indistinguishable from AI output.
  • Adverse selection can drive skilled providers out of such markets unless alternative, AI-resistant signals (liability, outcome guarantees) are available.
• Role of liability and institutions:
  • Liability contracts (verifiable damages) become the economically relevant mechanism to preserve market differentiation and sustain supply of high-competence providers.
  • Effective use of liability depends on enforceability (verification φ), litigation fixed costs (creating v_min), and insurance market design (retained risk vs pooled indemnity).
  • Policy levers: enhancing verifiability, lowering fixed procedural costs for small claims, and enabling credible enforceability can preserve professional-market functioning in the face of AI.
• Limits of provenance & disclosure:
  • Certifying human origin or mandating AI disclosure affects trust and perceived legitimacy but does not substitute for competence signals; provenance is not competence and cannot restore separation.
  • Mandatory AI disclosure may produce trust discounts (documented elsewhere) but does not solve information asymmetry about problem-solving competence.
• Design of insurance and guarantees:
  • Insurers can function as delegated verifiers/screeners; contract design (deductibles, experience rating) matters for whether liability preserves separation.
  • Policymakers and platform designers should consider how pricing of professional-liability insurance and claims-handling rules affect market signaling.
• Empirical priorities:
  • The proposed conjoint experiment targets the specific intersection (liability × AI disclosure × B2B expert services × WTP) currently missing in the literature; empirical validation is crucial.
  • Observational work should test predicted heterogeneity: sectors where value rests on verifiable outcomes should be less affected by artifact collapse than sectors where value is conveyed primarily by presentational artifacts.
• Caveats:
  • This is a theoretical and simulation study (preprint, not peer-reviewed). Simulation calibration is illustrative, not empirical proof.
  • The mechanism applies where buyers cannot verify outcomes directly; empirical heterogeneity (null findings in some contexts) is consistent with scope conditions.

If you want, I can: - Extract the formal propositions and equations (η*, φ·D ≥ v, κ_θ(D) formula) into a short math appendix. - Produce a one-page slide summarizing policy recommendations for regulators, professional associations, and insurers. - Draft the conjoint experimental protocol (attributes, levels, sample calculation) the author proposes.