Plan provides clarity for rights holders and AI developers, providing a clear mechanism for restitution for copyright violations

The second class of AI models are “large” models, which are the type of models that motivated the research in this paper. These models are trained on such a large volume of data that it is not possible to establish provenance ahead of time because, for the general operation of these models, no individual piece of content or even collection of content is significant enough to impact the vast majority of output from these AI models. This is not to say that the provision and availability of such AI models may not end up having a commercially harmful impact on a content provider. It is just that it is not clear in advance whether that might occur, nor is it clear what precise value specific content has in reducing AI training costs, even if the content used in training as a whole has an identifiable benefit. Once again, how content is used in training large models has been discussed considering human analogues. This is from OpenAI’s response to the New York Times’ lawsuit: Just as humans obtain a broad education to learn how to solve new problems, we want our AI models to observe the range of the world’s information, including from every language, culture, and industry. Because models learn from the enormous aggregate of human knowledge, any one sector—including news—is a tiny slice of overall training data, and any single data source—including The New York Times—is not significant for the model’s intended learning.5 Here is a scenario that captures the connection: Scenario 3a Teaching (AI Version): A developer builds a chatbot that teaches students economics based on a corpus of textbooks. Scenario 3b Teaching (Human Version): A person teaches students economics based on their reading of a corpus of textbooks. Note that in teaching, it is generally agreed upon that human teachers are not subject to copyright prohibitions, but it is precisely the issue with AI that remains an open copyright regulatory question. Once again, it is important to use the model to understand precisely why the human version involves a particular conclusion to see if it can assist in understanding the approach for the AI version. Analysing large AI models requires some adjustment to the baseline model of Section 2. First, the lack of provenance means that there is no opportunity for negotiation for the use of particular content, nor is it possible to know what the precise commercial harm (if any) to a content provider might be. Thus, the baseline model is amended to include uncertainty over potential commercial harm prior to setting prices for original content and AI use and to remove the ability to negotiate. Nonetheless, a copyright regime could prohibit (say, subject to large statutory infringement damages) the use of such content or permit its use. The model of Section 4 examines the expected social welfare outcomes from each of these traditional rights regimes and demonstrates that having no copyright protection is superior to copyright protection if the value of the content as a whole in lowering AI training costs is proportionately high relative to the expected commercial harm across all content providers. The large AI model context also creates an opportunity for a different type of rights regime. Section 4 proposes and analyses a mechanism that is an ex post ’fair use’ regime. Rather than intended to operate prior to potential infringement taking place, this mechanism operates ex post. This mechanism allows an AI model to be trained on all available content, and then, once original content owners observe leakage and realise commercial harm, exercising an option to sue the AI provider for lost profits. The regime sets a threshold on the magnitude of such harm, and those content providers who experience harm above that threshold exercise the option for lost profits while others do not. It is demonstrated that this mechanism offers stronger incentives for content providers than no copyright protection but also stronger incentives for AI training than traditional rights regimes. Moreover, subject to the financial feasibility of the AI provider, the threshold can be set low, implying that the mechanism will generate the highest expected social welfare amongst rights regimes.