Beaux-Arts Magazine – AI-proofing copyright

Last March, tens of thousands of illustrations in the style of Studio Ghibli, generated by ChatGPT, invaded social networks. These images caused quite a stir, ranging from homage to the mythical world of Japan’s Hayao Miyazaki to misappropriation of a genre unique to its author. This embodies the first side of a major legal challenge linked to artificial intelligence (AI) applied to art: the often unregulated use of protected works to train models (with data, known as “input”). Systems such as ChatGPT or Midjourney feed off huge corpora from the web, including works and content subject to copyright, exposing AI models and their pro- ductions to a risk of infringement. The 2019 European directive opens a loophole with its exceptions for “text and data mining (TDM)”, authorizing under certain conditions the use of protected works for the development of AI systems. Developers are thus free to engage in automated data extraction, except that rights holders are given the option of objecting via a special “opt-out” clause. However, the application of the opt-out to copyright remains complex. On the European side, a pioneering ruling by the Hamburg court extended the exception for scientific research, including the creation of a database used for AI training, rejecting the opt-out if the use is non-commercial.

United States divided

For the time being, rightful claimants can only hope to win their case if they can prove the reproduction or representation, even partial, of a work without authorization. A simple imitation “in the manner of” may not be sufficient to legally characterize an infringement in most countries. Under these conditions, it remains complicated for the various jurisdictions to convict the designers of AI systems, since the content generated, often inspired but not copied (and according to opaque technical modalities), escapes sanction, unless there is clear proof of reproduction of protected elements – even in the case of an ignored opt-out. Going in a different direction, the United States District Court in Delaware ruled that Ross Intelligence’s use of case law summaries from the Westlaw legal research site for the development of its AI violated copyright. By rejecting “fair use” (in the U.S., an exception to copyright allowing, in certain cases, the non-consensual use of a work), courts around the world could decide to favor stricter regulation under pressure from the cultural industries, forcing technology giants to rethink their model. Future rulings could trigger a paradigm shift by tightening regulations on such a global issue.