Generative AI Models at the Gate: Licensing frameworks for the effective and efficient protection of copyright protected content in an AI world

Expert economists Kadu Prasad and Jorge Padilla co-authored a paper, commissioned by the International Federation of the Phonographic Industry (“IFPI”), exploring licensing frameworks for the effective and efficient protection of copyright protected content in an AI world.

The views expressed in this paper are the sole responsibility of the author and cannot be attributed to Compass Lexecon or any other parties.

Executive Summary

A generative AI model is a type of artificial intelligence designed to create new content, such as text, images, music, or even code, based on patterns it has learned from existing data. These models are likely to have a material impact on all sorts of fields, including writing assistance, video generation, marketing, education, healthcare, gaming and finance.

Generative AI models are not designed to memorize data. They predict the next best word, pixel, or note based on learned probabilities. Then, when given a prompt, the models create responses that are coherent and contextually relevant. To derive those probabilities, generative AI models are trained on vast quantities of content – e.g., text, images, music, video, and/or code. While some of the data and content needed for their training can be freely used, some other is proprietary, and may be copyright protected and require prior authorisation. The widespread scraping of publicly-accessible data online raises concerns about the unauthorized use of copyrighted content for training AI and its impact on creative industries.

The core issue is, therefore, how the training of generative AI models interacts with copyright. Generative AI poses material threats to industries that rely on copyright protection, such as publishing, music, and film. AI-generated articles, books, and scripts are challenging traditional publishing by producing content at scale. The risk is that such AI-generated publishing content will flood the market, reducing the demand for human authors, and that publishers and news organizations will suffer from AI scraping and repurposing their copyrighted content without compensation. Likewise, because AI-generated music can mimic famous artists’ styles or produce convincing deepfakes, concerns have been raised about personality rights and copyright infringement in the music industry. In parallel, AI-generated scripts, deepfakes, and voice cloning are disrupting the audio-visual industries.

Copyright is a legal framework that grants creators exclusive rights to their original works, including literature, music, films, and visual art. It ensures that authors have control over how their work is reproduced, distributed, and monetized. Copyright law protects against unauthorized use, fostering a system where creativity is rewarded. We argue that the future of generative AI in the creative industries depends on protecting human creativity, which requires granting the rights that enable copyright holders to protect the content that generative AI models train on. Such protection is essential for the continuous development of valuable human-created content, which is essential for the development of valuable generative AI models. AI can serve as a powerful tool that enhances, rather than undermines, human artistic expression if copyright protection is secured.

Content creators (which, for the purposes of this paper, encompasses all categories of right holders – authors, artists, publishers and producers) have expressed concern over the unauthorised use of their content and have pursued legal actions for infringement of their intellectual property rights.[1] If copyright content is not, or only insufficiently, protected, there is a risk that those who create and invest in content may not obtain an appropriate remuneration for their investments. Generative AI developers could essentially ‘free ride’ on content creators’ investment, so that the incentive to produce content in the first place would falter. Ultimately, this not only could reduce the quantity and quality of human content directly consumed by firms and end consumers, but also could result in inferior generative AI models, which heavily rely on such content.

In response, policymakers across the globe have proposed regulations intended to ensure that generative AI models work well in the broad interest of society. Examples include the EU AI Act, [2] the proposed US Generative AI Disclosure Act, [3] and NO FAKES Act, [4] the Chinese Interim Measures for the Management of Generative AI Services, [5] etc. Some of these regulations are meant to foster transparency and accountability in AI development. The EU AI Act, for example, mandates that providers of generative AI models disclose information about the data used for training, including any copyright protected content, [6] which should enable right holders to protect the content that generative AI models train on.

Some appear to be concerned about a different risk, however; namely, that copyright could become a barrier for the development of new generative AI models. They fear that if copyright protected content, especially high-quality content, is not accessible in sufficient quantity and at a reasonable cost, generative AI developers, especially small start-ups, might not be able to rely on that content to train their models, which could result in inferior models with less valuable applications. In short, start-up providers of generative AI may be foreclosed.

Some of those so concerned have argued that training AI models should regarded as ‘fair use’, or outright exempted, so that it does not require any authorisation from copyright holders.[7] Others have proposed alternatives to market-based licensing by right holders, arguing for a downgrade of their exclusive rights in favour of statutory licences, levy mechanisms, compulsory collective management systems, or extended collective licensing.[8] In their view, the challenge is to design a copyright policy that offers enough protection to content creators, so that their incentives to invest in high quality content are maintained, without erecting a barrier for the development of new, high-quality generative AI models.[9]

In this paper, commissioned by the International Federation of the Phonographic Industry (“IFPI”), we assess whether there is any reason properly grounded in economics to depart from the status quo in the way copyright rules are applied to AI. Our answer is that the voluntary bilateral licensing of exclusive copyright rights is likely to maintain and promote the investment incentives of content creators and producers, as well as facilitate the development of better generative AI models.

The remainder of the paper is structured as follows. In section 2, we present a summary of our main conclusions. In section 3, we explain that copyright holders must be remunerated for the training of generative AI models. In section 4, we explain that copyright holders should have the right to deny the use of their protected content for the training of generative AI models. In section 5, we explain that the terms of use of copyright protected content for the training of generative AI models should be left to bilateral negotiations. In section 6, we rebut the arguments made in favour of the free use of copyright protected content for the training of generative AI models. In section 7, we offer a few concluding remarks. 

---

References