Darius Rostam (Bucerius Law School) has posted Opt-Outs in Copyright Law: Can They Save AI Training Under Fair Use and the Three-Step Test? on SSRN. Here is the abstract:
The use of copyright-protected works for AI training has thrust opt-out schemes into the center of copyright discourse. Unlike conventional copyright rules requiring prior authorization, opt-out schemes permit uses by default unless rights holders affirmatively object, effectively reversing copyright’s traditional logic of exclusivity. This Article provides the first comprehensive analysis of opt-out mechanisms in copyright law. Adopting a comparative perspective examining both US and EU copyright frameworks, it traces existing precedents ranging from copyright restoration to search engines, mass digitization projects, and online platform liability.
The Article argues that opt-out schemes serve two principal functions: preserving rights holder autonomy through flexible “loperty rules” that combine elements of property and liability rules, and controlling user infringement risk by forcing rights holders to disclose information necessary for transactions. However, these functions depend on two conditions: rightholders must be aware of both the use and their opt-out option, and the opt-out mechanism must be feasible with reasonable effort. Where these conditions are met, opt-outs can positively influence the fair use analysis, particularly factors one and four, by supporting an inference that a missing opt-out indicates no meaningful threat of substitution. Similarly, under the international three-step test, opt-outs help ensure compliance by preventing conflicts with normal exploitation and unreasonable prejudice to legitimate interests. Applying this framework to AI training under Article 4 DSM Directive and pending US litigation, this Article demonstrates that current opt-out regimes fail to meet these conditions. Rightholders often lack awareness, opt-out standards provide insufficient granularity, and many works are uploaded by third parties beyond rightholders’ control. Consequently, while opt-out schemes have legitimate doctrinal foundations, their application to generative AI training is currently dysfunctional.