Safe Harbor expansions can solve for infringing input AND output while also providing clarity necessary for these models to flourish

DMCA Safe Harbor. The Digital Millennium Copyright Act (“DMCA”) is a U.S. law created to address digital copyright issues that came about with the advancement of technology. The DMCA safe harbor provisions protect online service providers from legal responsibility for copyright infringement claims. DMCA protections might vary depending on a number of considerations, but we emphasize that they are not guaranteed for all model deployments. We examine several of these considerations here. DMCA protections for generative foundation models are uncertain. At first glance, it may seem like the Digital Millennium Copyright Act (DMCA) would protect machine learning model hosts, at least when users upload the model to be hosted (e.g., model repositories with a user interface, like HuggingFace). Like in other hosted sites, they would need to meet the relevant requirements like using a registered agent under DMCA §512(c)(2). Then they could put up a take-down request form and add filters for the offending model (or model output) when served with a take-down request under the DMCA §512(c) safe harbor.26 An internet company that has a notice-and-takedown scheme in place is not liable for hosting infringing content posted by a third party. But it is not obvious that the DMCA safe harbors apply to generated content. For example, in 2019, Amazon lost on a motion to dismiss when its algorithms selected copyrighted material to host on the website (Williams-Sonoma, Inc. v. Amazon.com, Inc., N.D. Cal. 2021, Order on Motion to Dismiss). The court was unconvinced that Amazon was eligible for safe harbor under the DMCA. They stated that to establish safe harbor the content must be stored “at the direction of the user” (Mavrix Photographs, LLC v. Livejournal Inc., 9th Cir. 2017, at 1052). This may mean that generated content does not have the same safe harbor and that post-hoc take-downs are not sufficient to reduce liability.27 As such, filtering of generated content before a takedown request is ever received may be more important while the courts determine the applicability of DMCA to generated content. It might matter where the model and data comes from. This also implies that DMCA protections may vary on who the model host and creator are. For example, a website hosting models uploaded by users might find it easier to argue for DMCA protection because the website itself is not creating or selecting the content (assuming that it follows other requirements like using a registered agent under DMCA §512(c)(2)). On the other hand, if a company were to create and host a model that itself selects content provided by others, like Amazon did in Williams Sonoma v. Amazon, it is unclear whether courts would agree that DMCA protections would apply. Another unclear variation on DMCA eligibility rests on the source of the training data: is the data user-contributed or creator-contributed? This might mean that, for example, DMCA safe harbors might be more likely to apply when a model is refined via Reinforcement Learning from Human Feedback (RLHF) (Ouyang et al., 2022) based on user ratings or updated automatically via user-generated data. These modes of method training and deployment are more akin to users uploading content to YouTube, and might help with arguments for DMCA protections. Things might become more murky if, for example, an RLHF model is trained on user data but this data is modified by creator-hired annotators. If model creators themselves curate and scrape data, then host the model themselves, this might be more akin to the Williams Sonoma case (which, again, is not a settled rule or a binding decision, but shows the variation of outcomes that is possible in the current state of the law). It is unclear what’s the best mechanism for DMCA takedowns with generative models. Even if it applies, it is unclear how the DMCA notice-and-takedown scheme would work as applied to foundation models, but the most likely “take down” approach might actually look more like output filtering with a safe harbor. As we will discuss in §4.2, instance unlearning is a nascent research area and retraining a model without a taken down datapoint could be exceedingly costly. The most likely approach in the near term is to “take down" model outputs that were too similar to a copyrighted work via a filtering mechanism since foundation models generate content on demand for users rather than hosting persistent content that can easily be taken down. But new research is needed to identify new and improved mechanisms for handling takedown requests in this relatively new setting. The Copyright Office has a DMCA anti-circumvention exemption for text and data mining. Finally, the Copyright Office also provides an exemption to DMCA’s anti-circumvention requirements in §1201 in the case of non-commercial data mining.28 This may allow non-commercial researchers to, say, remove digital rights management software to train on video or text content.29 Overall, DMCA protections are far from guaranteed, so model creators and deployers cannot rely on its safe harbor provisions to reduce liability. Instead, they must take a more proactive approach. Moreover, some previously-discussed provisions add to potential liabilities like §1202, creating additional compliance challenges