Copyrights are too costly --- liability rules are key

These cases are still ongoing, but the legal scholarship around them seems to suggest that the outputs of generative AI systems are unlikely to constitute infringement --- leaving the question of whether the remedies we grant for infringement when it does occur are commensurate with the harms suffered by artists & other creators. As mentioned earlier, fair use remedies are typically statutory damages or injunction, which in Sobel’s view “...overcompensate for the harms authors are likely to suffer from having their work included in training datasets.”72 Guido Calabresi and Douglas Melamed describe entitlements as being protected by the state in two ways: “property rules” or “liability rules.”73 “Property rules” which are currently used in the U.S. copyright system protect entitlement holders from [footnotes omitted] nonconsensual takings via injunctions or punitive damages, and are aimed at fully undoing takings.74 Such strong remedies enable entitlement holders to “set the price” of their own property, and are “premised on the assumption that the ideal allocation will occur if private parties bargain over resources.”75 Liability rules, on the other hand, permit an entitlement to be transferred in exchange for a fee set by a court or regulatory body using an “external, objective standard of value.”76 Mark Lemley and Philip Weiser point out that a major justification for liability rules is that transaction costs can impede efficient bargaining under property rule regimes, regardless of whether the property in question is real or intangible.77 Given the nature of AI development, where massive data sets are used and bargaining with countless individual entitlement holders would be incredibly costly, some have argued that the adoption of liability rules may be an effective way to balance the equities, since this would promote AI development while compensating entitlement holders in a standardized way. 78