Neither laissez-faire nor blanket IP protects the future of creative content for AI: permissive access starves creators of compensation, while strong rights blunt originality — and even a high‑quality model can erode itself by causing creators to homogenize their output. A data intermediary that internalizes cross-creator externalities and subsidizes novel contributions can both preserve incentives and improve long-run model quality.

Main Finding

The paper shows that the two prevailing policy extremes for sourcing human-created content to train AI—(1) an unrestricted “free-for-all” (fair use) and (2) a strict individual intellectual-property (IP) rights regime—both fail to produce socially desirable outcomes. Free-for-all destroys creators’ incentives. Surprisingly, strict IP also under-provisions original creative effort because AI firms’ monopsonistic buying power and correlation across creators’ output depress prices and impose an “originality penalty” that disproportionately discourages novel creators. In a dynamic extension, a high‑quality model causes humans to rely more on AI, which homogenizes future human-created data and feeds back to degrade model performance—a “curse of precision.” The authors propose a data intermediary that bundles creators, internalizes cross‑creator externalities, and distributes payments (using Aumann–Shapley–inspired weights) via simple two‑part tariffs to restore efficiency (under full information).

Key Points

• Free-for-all (blanket fair use): Leads to underinvestment in human-created content because creators are uncompensated while AI outputs substitute for original content.
• Strict IP / individual bargaining: Fails even without transaction costs because
  • AI firms have monopsony power and optimally buy lower‑quality (cheaper) content, reducing creators’ incentives.
  • Correlation across creators means selling one creator’s data partially reveals others’ content, creating negative externalities that depress the marginal price creators receive.
  • Originality penalty: Highly novel/innovative creators are undercompensated relatively more than typical creators because AI prefers representative data; scarcity of originality does not command a premium in equilibrium.
• Dynamic "curse of precision":
  • Better models substitute for costly human creativity, increasing human reliance on AI-assisted generation.
  • This increases homogeneity of human-produced data, which is then used for retraining, degrading the model over time—even if fresh human data continues to be supplied.
  • This is a distinct economic pathway to phenomena similar to empirically observed “model collapse.”
• Data intermediary solution:
  • A single intermediary negotiates with the AI firm, pooling creators to overcome monopsony power and internalize information‑leakage externalities.
  • Payments allocated according to marginal contributions (weights inspired by Aumann–Shapley) can reverse the originality penalty.
  • A two‑part tariff (affine lump-sum and affine per-effort transfers) is sufficient to restore efficiency under full information.
• The proposed intermediary is a benchmark: it solves the market failures in a full‑information setting but does not yet address informational asymmetries or agency frictions, which are left for future work.

Data & Methods

• Modeling approach:
  • Static model: Stackelberg game where an AI firm (buyer/monopsonist) first chooses price(s) for content/quality and individual creators choose effort/quality. Analysis focuses on equilibrium effort vs socially optimal effort.
  • Dynamic model: Continuous‑time framework where the AI firm repeatedly trains on human-generated data; humans in turn use AI to assist creation. The model captures feedback effects between model quality and human effort/heterogeneity over time.
• Key assumptions:
  • Creators expend costly effort to produce content; content signals are statistically correlated across creators (substitutability).
  • AI firm has market power (monopsony/oligopsony) in acquiring training data.
  • Full information in the intermediary benchmark (no asymmetric information or principal‑agent frictions).
  • Social planner objective compares aggregate surplus including human production value and AI output.
• Analytical tools:
  • Game-theoretic equilibrium analysis (Stackelberg and continuous‑time dynamics).
  • Comparative static analysis to identify originality penalty and underinvestment.
  • Cooperative-division ideas (Aumann–Shapley style weighting) to apportion payments among creators in the intermediary solution.
  • Formal propositions and proofs provided in the Appendix.
• Literature connections:
  • Extends informational‑economics literature on correlated data and data markets (Acemoglu et al., Bergemann et al.).
  • Links to ML literature on model collapse and self‑training dynamics (Alemohammad et al., Shumailov et al.).
  • Situates results in IP policy and fair-use legal debates and ongoing litigation.

Implications for AI Economics

• Policy design:
  • Neither blanket fair use nor uncompensated individual IP bargaining is sufficient; policy should enable compensation mechanisms that internalize cross‑creator externalities and mitigate AI monopsony power.
  • Data intermediaries (or collective bargaining platforms) are promising institutional solutions to align incentives between creators and AI firms.
  • Regulatory attention should target market structure (monopsony power of AI buyers) and mechanisms for collective payments and attribution.
• Market design:
  • Markets for training data should reward originality explicitly (reverse the originality penalty) to preserve diversity and innovation in creative inputs.
  • Simple two‑part tariffs can implement efficient transfers under full information; practical implementations must also handle informational asymmetries and agency problems.
• Dynamic risk management:
  • Economic incentives, not only technical fixes, are central to preventing deleterious feedback loops (homogenization → degraded models).
  • Platforms and regulators should monitor and counteract incentives that push creators toward homogenized, AI‑assisted outputs (e.g., by subsidizing novel content or limiting overreliance on model outputs in training pipelines).
• Research directions:
  • Study intermediary designs under asymmetric information and principal–agent constraints.
  • Empirically measure content correlation, originality value, and the strength of monopsony power in data markets.
  • Explore hybrid legal/regulatory approaches combining clearer property rules, mandated intermediaries, or antitrust remedies to monopsonistic data buying.
  • Investigate platform-level mechanisms (recommendation systems, attribution, payment rules) that preserve content heterogeneity and long‑run model health.

Overall, the paper reframes the AI-data policy debate from a binary copyright question to a market‑design problem: effective institutions (not just stronger IP or unfettered access) are required to sustain both technological progress and human creative incentives.