Safe Harbors for AI models provide the necessary policy backing to allow the industry to continue to grow while strengthening copyright protection

As demonstrated throughout this work, the risk of copyright violation and litigation, even with fair use protection, is a real concern. To mitigate these risks, we recommend that foundation model practitioners consider implementing the mitigation strategies outlined here and pursuing other novel research in this area. There is significant, exciting, technical research required to make technical mitigation strategies robust and aligned with fair use doctrine. We reinforce that machine learning researchers must play a role in providing viable mitigation mechanisms to demonstrate that models are truly covered by fair use. Preventing extreme outcomes in the evolution of fair use law by advancing mitigation strategies. Legal scholars have noted that there might be two extreme outcomes for fair use and machine learning (Sobel, 2017). On one hand, there is a possibility that courts may rule that foundation models are widely acceptable under fair use regardless of the likelihood of infringement or efforts at mitigation, which could have adverse effects on the income of data creators and disregard the ethical and moral rights attached to their work. On the other hand, there is a possibility that courts may declare that generative foundation models cannot be trained on unlicensed copyrighted data in most cases. This scenario could lead to a concentration of power for companies that have retained licenses to large amounts of data; companies like YouTube or Facebook might be able to leverage large amounts of user-contributed data where others would be shut out of model training. Neither of these two outcomes is ideal. As litigation progresses, identifying mechanisms to prevent extreme outcomes will be critical. For example, it is important to understand that not all forms of foundation models are inherently infringing and that some of their potential risks can be effectively managed through technical means. With better demonstrations of co-evolving technical mitigation strategies, the law might find a middle ground that allows model training and deployment with sufficient effort to implement objectively strong mitigation strategies. Courts may consider the reasonable efforts of model builders and deployers to mitigate copyright risk, both in deciding fair use and in determining whether they can face indirect infringement liability. Trademark courts have taken a similar approach, for example in Tiffany v. eBay (2d Cir. 2010). As such, advancing research in this area (with methods such as improved similarity metrics) may help in preventing extreme outcomes in legal settings.42 We should not over-zealously filter. There must be a balance to filtering. Well intentioned but strict filtering mandates adopted by other countries have been criticized and criticized for their impacts on free speech (Schmon et al., 2022). Similarly, YouTube’s content ID system, a large-scale filtering approach, has been criticized for not following fair use standards and being overaggressive in its filtering (Bartholomew, 2014; Boroughf, 2015). Levendowski (2018) points out that restrictive views of fair use doctrine can exacerbate biases and that fair use can help create fairer systems. While mitigation strategies will help prevent undesirable outcomes, it is important to develop strategies that carefully align with fair use standards, as we have previously discussed. This means that factual content should not necessarily be filtered, neither should parodies, or short form regurgitation used for commentary. And evolutions of fair use doctrine or further policymaking should consider the distributive effects of preventing access to certain types of data for model creation. Policymakers could consider how and if DMCA (or similar) safe harbors should apply to foundation models. As we have seen, there are various ways, including filtering, to mitigate the risk of copyright infringement in the output of foundation models, but none will entirely eliminate the risk of liability. Even when trained on presumably permissively licensed datasets, for example, it is difficult (if not impossible) to determine the provenance of every piece of data and filter it out. Users might post content to seemingly permissively-licensed databases that they do not actually have the rights for. There may even be uncertainty about whether a piece of content is actually in the public domain or whether that status has been revoked.43 And even if foundation model practitioners implement strong mitigation strategies, the amorphous nature of fair use doctrine may make it difficult to know what kinds of content will be covered by fair use ex ante. With the uncertainty of DMCA protections (discussed in § 3), the law may need to adapt to this reality, and it could do so, for instance, by clarifying the role of safe harbors for models that implement sufficiently strong mitigation strategies. Policymakers could make clear that DMCA protections apply to this setting or they could identify other more suitable safe harbor mechanisms. This may provide more balance than general-purpose text and data mining exemptions seen in other countries, but again are not a panacea. Such safe harbors would have to be structured to consider the strength of the implemented mitigation strategies to ensure that they are not abused.