Copyright Extremes Undercut AI Training Markets

AI training depends on human-generated content, but the paper argues that the policy debate has been framed too narrowly. A free-use regime gives AI developers abundant data while leaving creators uncompensated. A strong-property regime gives creators bargaining power, but the authors show that it still fails to support the right kind of content creation. The central claim is that AI data markets need mechanism design, not a binary choice between fair use and exclusion rights.

Core Contribution

The paper contributes a formal market-design argument for AI training data. In the static model, creators decide how much costly effort to invest in content quality before an AI firm trains on the resulting content and sells model output. The firm benefits from better content, but it does not internalize the full incentive problem faced by creators. Even when creators hold strong intellectual property rights and can be paid for training use, the equilibrium underprovides creative effort.

The sharpest result is the paper’s “originality penalty.” More original creators are exactly the contributors whose work is less substitutable and more useful for improving the model, yet the market structure gives them weaker incentives than efficiency would require. The reason is not that the firm ignores their value; it is that bilateral compensation after content has been produced cannot fully reward the ex ante effort that made the content valuable. The inefficiency is therefore built into timing and bargaining, not just into low payment levels.

Technical Approach

The static analysis is framed as a Stackelberg game. Creators move first by choosing content-creation effort. The AI firm then observes the resulting content environment and chooses how to train or contract under different rights regimes. The efficient benchmark asks what effort levels would maximize total surplus if the downstream training benefits and upstream creation costs were coordinated.

Against that benchmark, the free-for-all regime fails in the expected direction: creators supply too little effort because training use generates no direct compensation. The stronger result is that strong IP rights also fail. Granting exclusion rights changes the transfer problem, but it does not remove the gap between the social value of original content and the creator’s private return from producing it. In the paper’s terms, originality can reduce incentives rather than increase them.

The dynamic extension adds a feedback channel between AI performance and future human content. A more precise AI model makes AI-assisted creation more attractive. Humans then rely more on AI when producing new content, which makes future training data more homogeneous. The authors call this the “curse of precision”: better model quality can induce a training-data shift that later degrades model performance. The mechanism is economic rather than purely statistical. Model improvement changes human production behavior, and that behavioral response changes the data distribution.

Market Design Proposal

The proposed fix is a data intermediary that bargains collectively with the AI firm and uses the proceeds to subsidize innovative contributions. Collective negotiation addresses the firm-side contracting problem, while targeted subsidies address the creator-side effort problem. The intermediary is not presented as a generic data broker; its role is to reshape incentives so that original content is produced before the AI firm trains.

The design also separates compensation from simple usage counts. If payments only track how much content is used, they can reward common or easily substituted material while missing the margin that matters for model improvement. The paper’s subsidy logic instead points toward rewarding contribution to diversity, originality, and future training value.

Results and Analysis

The evidence is theoretical: propositions, comparative statics, and equilibrium comparisons rather than measurements on an empirical dataset. That is appropriate for the question the paper asks, but it changes how the result should be read. The paper does not estimate the size of the originality penalty in an existing market, and it does not test whether a real intermediary can measure innovative contribution well enough to implement the proposed subsidies.

Within the model, the contribution is clean. The paper shows that 2 commonly discussed regimes fail for different reasons, identifies 2 distinct failure modes, and proposes a mechanism that targets both. The static result is useful because it challenges the assumption that strong IP alone solves creator incentives. The dynamic result is useful because it connects training-data policy to future model quality: a good model can make its own next training set worse by shifting human production toward AI-assisted sameness.

The practical implication is narrower than a policy prescription but still valuable. For AI firms, licensing content without changing upstream incentives may buy legal access without preserving the data frontier. For creators, stronger rights may improve bargaining position while still leaving original work under-rewarded. For policymakers, the paper argues that the relevant design question is how to finance and allocate rewards for high-value human originality before it disappears from the training pipeline.