The rise of generative artificial intelligence (AI) has sparked significant concerns among content creators, including editors, translators, illustrators, and voice actors. These AI models rely on large databases of diverse documents to function, undergoing a training phase where algorithms identify patterns in the data. Many content creators find it unjust that companies such as OpenAI (the developer of ChatGPT and DallE), Anthropic (creator of Claude), and Microsoft (developer of Copilot) profit from their work without paying royalties.
In the United States, multiple class action lawsuits are underway, aimed at addressing these issues. A potential outcome could result in significant financial reparations for creators, as highlighted by Anthropic's $1.5 billion out-of-court settlement with a group of authors.
Egeda) in collaboration with Red.es.
Protecting Intellectual Property Rights in the AI Era
Q. Is it possible to protect intellectual property rights in the age of AI?
A. This question has a complex answer. Intellectual property rights are fundamentally preventive; if you intend to utilize a resource for training AI, you either obtain permission from the rights holders or rely on exceptions that allow usage without permission. Transparency is another critical issue: knowing which data has been used in AI training is essential. The European Union mandates transparency through its AI Regulation, though jurisdictions like the UK are still exploring the necessity of such measures.
In France, a new bill proposes a presumption of use, where it is assumed that content has been used for training unless the AI developer can prove otherwise. This paradigm shift would ease legal enforcement, contrasting with the typical requirement that the copyright owner must demonstrate infringement.
Regulatory Framework and Fair Use
Q. Are there exceptions in EU regulation that facilitate data set usage? Do you think European regulation offers sufficient protection?
A. While the EU has text mining exceptions, they come with strict requirements. These exceptions generally pertain to specific activities, primarily non-profit research. Additionally, they hinge on the legality of content access, meaning that only legally obtained content can invoke exceptions. This aligns with the ‘fair use' doctrines in the United States, creating uncertainties around the interpretation of such provisions.
Q. Does the argument of fair use apply equally to AI training as it does to traditional learning methods?
A. This is a longstanding debate, intensified since the introduction of the EU's text and data mining exceptions. While it's conceivable for a person to read and learn from many books, an AI model ingests large volumes of content, involving acts of reproduction reserved for copyright holders. A clear comparison between human learning and AI training is challenging.
Current discussions also focus on the memorization of training documents by AI models, with studies suggesting a higher level of memorization than previously acknowledged. Legally, the distinction between memorization and immediate reproduction is irrelevant; what matters is the original source and whether there's an act of reproduction involved.
The Impact of Lawsuits and AI Development
Q. How do illegal sources factor into these discussions?
A. A judge's ruling led Anthropic to settle with authors, emphasizing that content from pirated sources cannot invoke ‘fair use.' Numerous related cases in the US are still pending, and existing reports indicate that fair use defenses will not apply if the content source is illegitimate.
Q. Are AI developers facing greater challenges due to these rulings?
A. Yes, developers must diligently verify source legality, as training on illegal content could compromise the entire process. Caution is crucial.
Q. Will we see more out-of-court settlements similar to Anthropic's in the United States?
A. It's possible, given the high costs associated with litigation in the US. In Europe, while some agreements may emerge, it is less likely. Future litigation will likely focus on text mining exceptions and liability in result generation.
Broader Concerns and Future Directions
Q. What about other forms of content, such as artists, voice recordings, and music? Are they facing similar legal challenges?
A. Yes, the situation is concerning in Europe as well, with celebrities exploring ways to protect themselves, such as registering trademarks for their vocal performances, similar to initiatives by Taylor Swift and Matthew McConaughey. Various legal frameworks—including personality rights, trademarks, and GDPR—apply here. Denmark has even proposed new digital rights related to replication.
Q. Do you think the French presumption of use model will influence European legislation?
A. Some advocate for such changes, especially in light of potential revisions to the Intellectual Property Rights Enforcement Directive 2004. Implementing the French model involves reconciling diverse member state interests.
Q. What does the future hold for European legislation concerning property rights and AI in the next decade?
A. I expect a more developed legal framework along with increased jurisprudence. It is crucial not to assume that new technology negates existing laws. We should avoid overly specific laws that may quickly become obsolete, ensuring flexibility to adapt to future developments. The continued global nature of AI challenges us to form international strategies that include minimum safeguards and protections.
