Win-win: How to Remove Copyright Obstacles to AI Training While Ensuring Author Remuneration (and Why the European AI Act Fails to Do the Magic)

Abstract

In the debate on AI training and copyright, the focus is often on the use of protected works during the AI training phase (input perspective). To reconcile training objectives with authors' fair remuneration interest, however, it is advisable to adopt an output perspective and focus on literary and artistic productions generated by fully-trained AI systems that are offered in the marketplace. Implementing output-based remuneration systems, lawmakers can establish a legal framework that supports the development of unbiased, high quality AI models while, at the same time, ensuring that authors receive a fair remuneration for the use of literary and artistic works for AI training purposes – a fair remuneration that softens displacement effects in the market for literary and artistic creations where human authors face shrinking market share and loss of income. Instead of imposing payment obligations and administrative burdens on AI developers during the AI training phase, output-based remuneration systems offer the chance of giving AI trainers far-reaching freedom. Without exposing AI developers to heavy administrative and financial burdens, lawmakers can permit the use of the full spectrum of human literary and artistic resources. Once fully developed AI systems are brought to the market, however, providers of these systems are obliged to compensate authors for the unbridled freedom to use human creations during the AI training phase and displacement effects caused by AI systems that are capable of mimicking human literary and artistic works.

As the analysis shows, the input-based remuneration approach in the EU – with rights reservations and complex transparency rules blocking access to AI training resources – is likely to reduce the attractiveness of the EU as a region for AI development. Moreover, the regulatory barriers posed by EU copyright law and the AI Act may marginalize the messages and values conveyed by European cultural expressions in AI training datasets and AI output. Considering the legal and practical difficulties resulting from the EU approach, lawmakers in other regions should refrain from following the EU model. As an alternative, they should explore output-based remuneration mechanisms. In contrast to the burdensome EU system that requires the payment of remuneration for access to human AI training resources, an output-based approach does not weaken the position of the domestic high-tech sector: AI developers are free to use human creations as training material. Once fully developed AI systems are offered in the marketplace, all providers of AI systems capable of producing literary and artistic output are subject to the same payment obligation and remuneration scheme – regardless of whether they are local or foreign companies. The advantages of this alternative approach are evident. Offering broad freedom to use human creations for AI training, an output-based approach is conducive to AI development. It also bans the risk of marginalizing the messages and values conveyed by a country’s literary and artistic expressions.

Keywords: copyright, text and data mining, freedom of expression, art autonomy, reservation of rights, three-step test, domaine public payant, equitable remuneration, levy system, collective rights management, trustworthy AI, right to research, unbiased AI