Plan protects the rights of content creators and promotes responsible AI development while also setting the stage for dynamic AI policies going forward

Both of the traditional rights regimes involve inefficiencies: (CP) potentially restricting the use of original content and both regimes trading off the incentives of the OCs and AI. The question is whether a different mechanism can create a more favourable balance of incentives for both original content investment and AI training quality while preserving the use of original content in consumption and AI training. Recall that, in the small AI model context, the AI provider would agree not to use content if the potential harm from leakage of that content were high. Of course, here, the extent to which leakage might be harmful to individual content providers is not known ex ante. It is, however, known ex post. This suggests a potential mechanism that involves liability assessed ex post is feasible. That mechanism would proceed as follows: 1. If ρi(1) < Γ, the AI provider is not liable for any damages from the use of i’s content. 2. If ρi(1) ≥ Γ, the AI provider must pay full compensation to i of an amount D(ρi(1)) = 1 4ρi(1)ˆxFU Here (FU) stands for ‘fair use’, and ˆxFU is the chosen content quality under that regime. The level of damages, if triggered, allows the content provider to earn 1 4 ˆxFU as if the leakage never occurred.11 The following proposition characterises the outcomes under (FU) and compares it to traditional rights regimes. Proposition 4 Suppose that AI earns non-negative profits under the (FU) mechanism. In the large AI model, for each OC, i, ˆxi,CP ≥ ˆxi,FU ≥ ˆxi,NC and for AI, ˆyFU ≥ ˆyCP ≤ ˆyNC. Using the specific function forms, expected social welfare under (FU) will exceed that under (CP) if and only if: γ > E[ρ(1)|Γ]2 + E[ρ(1)|Γ] − E[ρ(1)](1 − E[ρ(1)|Γ]) 4(1 − E[ρ(1)|Γ]) where E[ρ(1)|Γ] = � Γ 0 ρi(1)dF(ρi(1)). The proof (in the appendix) compares the first-order conditions for each type of investment and follows on by calculating social welfare under (FU) and compares it to those under traditional rights regimes. The condition on AI profits arises because it is possible that Γ is such that: u(ˆyFU) − cAI(ˆyFU, ˆxi,FU) < (1 − F(Γ)) � 1 Γ 1 4ρ(1)ˆxi,FUdF(ρ(1)) in which case AI would not be finally viable because expected damages are above its profits from AI. Below, when we examine what Γ is chosen by the planner, this feasibility constraint needs to be satisfied. The (FU) mechanism has two advantages over traditional rights regimes. First, under the (FU) mechanism, AI chooses to use all content in training the AI (i.e., s = 1).12 Thus, the prices for content and the AI have the same structure as under (NC) but, importantly, in terms of the consumption of that content, all consumers will do so (potentially) with half of the market purchasing from OC and the other half relying on the AI product. Thus, the (FU) mechanism allows content to be used at its socially desirable level for AI training and also to be consumed relatively widely. Second, this has an impact on incentives. As all content is used in AI training, AI training quality under (FU) is the same as that under (NC) for fixed content quality and, therefore, highest among the three regimes compared as OC content has a higher quality under (FU) than (NC). OC content is higher than under (NC) because if potential harm from AI training exceeds a certain threshold, each OC expects to receive a damages payout. Thus, from their perspective, while under (NC) their potential appropriation is 1 4(1 − E[ρi(1)]) for each unit of xi, under (FC) it is F(Γ) 1 4 � 1 − E[ρi(1)|Γ] � + (1 − F(Γ)) 1 4, a higher marginal return as E[ρi(1)|Γ] ≤ E[ρi(1)] (with an equality if Γ = 0). These advantages hold for any Γ < 1. But what level of Γ would be optimal? Note that expected social welfare is higher as Γ is reduced, and there is a lower threshold of harm that can trigger compensation. This has the effect of raising both content provider incentives and AI training incentives. The latter effect arises because, regardless of Γ, all content is available for AI training. Moreover, a lower Γ does not change the total number of consumers who consume original content (that is, 1 2 + 1 2E[ρ(1)]). Note that as Γ → 0, ˆxFU → ˆxCP and ˆyFU → 1 + γ 1 4. This is feasible if AI profits (u(ˆyFU) − cAI(ˆyFU, ˆxFU) − � 1 0 1 4ρ(1)ˆxFUdF(ρ(1))) remain positive. Compared with (CP), Γ = 0 behaves like a regime where all content is copy-protected. However, it is permissive of infringement with compensation, which generates the benefits of original content for AI training without diminishing the incentives of original content providers to invest in quality. In effect, the difference is akin to a prohibition versus a Pigouvian-like tax on the use of original content where external effects are internalised through ex-post prices. In practice, an (FU) mechanism is respectful of transaction costs associated with asserting copyright protection when limited harm occurs. For this reason, it also rationalises the human scenarios explored earlier in that those scenarios are very unlikely to cause commercial harm to original content providers, even in the aggregate. 5 Conclusion There is a sense in which the copyright challenges associated with generative AI are not that different to those that arose with digitisation and the rise of the Internet. In each case, a new technology lowered the cost of utilising original content while enhancing its potential scope or value. For generative AI, this arises because such AI can be trained on original content but also, precisely because that content is widely available, reproduce that content within AI products themselves. This threatens existing business models for original content providers that, not surprisingly, they are protective of. Copyright law played a role in the evolution of digital technologies. However, one thing that was preserved was long-standing copyright protections, including the ability to prevent use except where that use was regarded as fair. This paper has analysed generative AI, taking into account the particular role of such technologies as well as how generative AI products threaten original content creators’ commercial activities. It is demonstrated that different rights regimes can lead to distinct social welfare outcomes, but these rest firmly on the ability of original content providers and AI providers to negotiate over the use of copy-protected material in AI training. When they can negotiate – something that requires knowledge of what content is proposed to be used in AI training as well as what potential harm to original content providers’ commercial activities it might cause – then traditional full copyright protection leads to stronger incentives to produce original content and that this determined the social welfare properties of such protections. The conditions that allow for this, however, require that the AI models be small in the sense of using and relying on a relatively small corpus of content. Such “fine-tuned” or RAG models are currently being developed. The value of copyright protection, however, rests on the potential for commercial harm. If that potential is low, as it arguably is for similar human uses of content in their own “training,” then a strong protective regime may not be of much social value. Consequently, when the AI models are large in the sense of using an almost unimaginable amount of content in training, the type of rights regime imposed can have important social welfare consequences. Because negotiations are prohibitive as it is hard to identify copyright use, let alone determine clearly whether copy-protected content has been used in training, copyright protection, while potentially strengthening incentives to create original content, serves to limit its use both for consumption and training, which involves adverse social consequences. No copyright protection does not involve those costs, but if the threat to original content providers’ commercial activities is high, it could undermine the production of original content and, in the process, its uses. However, if the threat to those commercial activities is low, as it arguably is for equivalent human uses of content, then a permissive regime for the use of content in AI training is socially desirable. For that reason, this paper identifies an alternative mechanism based on fair use. That mechanism permits the use of copy-protected content in AI training but subjects AI providers to damages should the realised commercial harm of content providers be large. Thus, the purpose here is to provide some insurance. The presumption is that content used in AI training is, for the most part, not likely to damage original content providers’ commercial interests. However, should it turn out to do so, they will receive compensation. Insured against such risks, the original content providers’ incentives to create content are undiminished. This represents a practical way of respecting copyright ownership while also allowing its large-scale use in AI training in a way that minimises transaction costs that might arise from injunctive legal action or license negotiations. It is also conceivable that copyright holders could opt out of this type of regime, although the technical requirements to do so would likely require further development as it would require content used to be traceable. When a new technology arises that creates copyright challenges, a clarification of the rights regime can often lead to institutional and technological developments to “make it work.” This happened for music rights with radio and public broadcasting in the form of collecting societies. For AI, Besen (2023) argues that the large volume of data lends itself towards collecting copyright payments by collective societies as seen in other domains such as music. He believes that this would have to be established by government regulation. However, in other areas, the developments were technological. For instance, Google developed a rights management system for YouTube that allowed people to post content with copy-protected materials but for rights holders to be notified and then to utilise a revenuesharing arrangement (Gans (2015)). By settling on a policy approach to generative AI, even if some inefficiencies remain at present, these may incentivise new ways of minimising those inefficiencies.