Permission to Learn: An Economic Approach to AI Copyright

Review

Joshua Gans argues that general-purpose AI training requires a copyright arrangement that does not depend on negotiating permission for each work in advance. In this policy essay, drawing on his 2026 *Journal of Law and Economics* article, he distinguishes the economics of training on vast corpora from licensing identifiable sources used in retrieval. With millions of works, uncertain marginal contributions and harms observable only after deployment, he contends that an advance permission regime can suppress beneficial training without reliably producing a functioning market. He instead proposes an ex post, fair-use-like mechanism permitting general training and compensating creators when outputs demonstrably substitute for protected expression.

The argument is more discriminating than an assertion that AI should be free to learn. Sections 3 and 4 compare unrestricted training and full copyright control before proposing a liability rule under which harmed creators recover lost profits above a threshold, rather than statutory damages attached to training copies. Section 5 distinguishes expressive substitution from lost sales caused by competition over facts or ideas. It also cautions that contemporary licensing contracts, often bundling training with access, display rights or other services, cannot retrospectively prove that licensing was feasible when earlier models were trained. Retrieval of identifiable news or other current material is explicitly treated as a case where bargaining may work. Gans also preserves a role for opt-out rights when a creator's objection is moral rather than economic.

This proposal makes a governance trade explicit. Permission allocates an initial exclusion right to creators; retrospective compensation shifts the training decision to AI providers and turns protection into an enforceable claim arising after deployment. That can reduce holdout and clearance costs, including barriers faced by smaller model developers. The associated loss of creator control is not simply a side effect: it is the mechanism's institutional premise. The argument's welfare logic has force precisely because universal advance negotiation may be impossible, and Gans acknowledges both administrative claims costs and the risk that large firms are better positioned to transact or litigate.

The essay does not, however, establish the institutional machinery necessary to make its proposed insurance analogy reliable. A claimant must discover relevant outputs, establish expressive substitution and causation, reconstruct profits in a counterfactual world without the challenged AI use, and bring a claim whose expected value exceeds legal and evidentiary costs. The provider may hold the most useful training, deployment and output evidence. Gans acknowledges measurement difficulty and small-creator disadvantage but does not specify evidence-access obligations, collective claims, independent assessment, proportional claims procedures or funding arrangements. A damages threshold economises on administration while deliberately leaving some losses uncompensated. This is an implementation dependency, not a reason to dismiss the transaction-cost diagnosis.

A further distinction matters for the paper's reach. The proposed measure targets protected expressive substitution, not every livelihood impact of generative AI. It does not supply a general remedy for market dilution, reduced bargaining power, or declining returns from competition with outputs that do not reproduce protected expression. Nor does this short essay empirically validate a threshold, estimate claims-system costs, or demonstrate that net gains persist under concentrated control of evidentiary infrastructure. Those questions require analysis beyond the essay and should not be attributed to its underlying scholarly model without examining that model independently.

The policy judgment is consequently conditional. Liability rules may be preferable where advance licensing is genuinely infeasible and harms are sufficiently observable, attributable and remediable. Their legitimacy in operation depends on whether the institutions administering compensation can prevent the evidentiary and litigation burden from extinguishing the very claims the rule creates. A credible next step would test representative creator claims across different market sizes, with defined disclosure duties, causal standards, affordable adjudication and measured recovery rates. That evidence would determine whether the regime preserves creator incentives in practice rather than only in theory.

Key Insight

Replacing permission with compensation does not remove copyright governance costs; it shifts control over training to AI providers and makes creators' protection depend on access to evidence and enforceable redress.