Use of copyright works in artificial intelligence training

The use of protected works to train artificial intelligence is at the centre of one of the biggest legal controversies of the decade. A recent US ruling, Kadrey v. Meta, has provided new clues as to how these disputes might be resolved.

Context

In recent weeks, two relevant rulings have been handed down in the US on the use of copyrighted works in the training of artificial intelligence models. One of them is the case Kadrey v. Meta Platforms Inc. where a group of thirteen authors, mostly fiction authors, sued Meta (the US technology company that owns platforms such as Facebook, Instagram and WhatsApp, currently one of the leading companies in the development of artificial intelligence and virtual reality) alleging copyright infringement for the unauthorised download and use of their books available online to train Meta's generative AI models (specifically, their large language models, called Llama, Large Language Model Meta AI).

What initially appeared to be a clear copyright infringement, as Meta was found to have used datasets containing millions of books and academic articles downloaded from illegal sites, was eventually resolved in favour of the technology company under the "fair use" exception. In US law, "fair use" is an exception that allows the use of copyrighted works without the permission of the copyright holder in specific cases, provided that certain criteria are met according to four main factors:

1. purpose and character of the use (whether it is commercial or educational, and whether it transforms the original work); 2. nature of the copyrighted work.

2. Nature of the copyrighted work.

3. Amount and substantiality of the use in relation to the work as a whole.

4. Impact on the market for the original work or its economic value.

The court's decision

The Northern District Court of California decided to uphold Meta's defence, based on the following key points:

1. Transformative nature of the use (Fair Use Factor 1):

The court held that there was no doubt that the use of the books to train an AI model had a "highly transformative" purpose and character, distinct from that of the original works.

2. Market harm (Fair Use Factor 4):

The decision focused on the plaintiffs' failure to demonstrate concrete harm to the market for their books. The judge emphasised that it was not sufficient to

The judge emphasised that it was not enough to speculate that the AI model might replace authors or reduce licensing opportunities, but that it was necessary to provide specific and solid evidence, which was not forthcoming.

3. Limitations of the judgment:

The Court emphasised that its decision does not imply that the use of copyrighted works to train AI models is lawful in general. It is a ruling limited to this group of plaintiffs, who - in the court's view - failed to properly structure their arguments or present compelling evidence. The judge even hinted that, in other cases, with better grounds, it could be ruled that this type of use does not constitute fair use.

Implications for rights holders and companies

The ruling is a procedural victory for Meta, but it does not definitively settle the controversy. In fact, the court itself acknowledged that AI training with unauthorised protected works may not be legal in many cases.

At the same time, the ruling sends a clear signal: litigation in this area will depend to a large extent on how the arguments are framed and the type of evidence presented, especially with regard to the economic impact on the markets for original works.

Moreover, the case is part of an evolving and still uncertain regulatory landscape. In May, the US Copyright Office published a preliminary report in which it concluded that AI training with protected works, without permission, probably does not constitute fair use (the publication of this report was followed by relevant changes in the direction of the US Copyright Office).

Recommendations

For technology companies, publishers, authors and rights holders in Chile and the region, this case offers several lessons:

-Due diligence: verify the provenance and licensing of datasets used to train AI models.

-Contractual management: enforce clauses on use of works in licensing, publishing contracts or collaboration agreements with technology companies.

-International monitoring: closely follow court decisions in the US and Europe, as they will establish relevant criteria for future disputes.

-Prevention and legal strategy: develop internal policies to protect copyrights against unauthorised use, or to legally shield AI training processes.

How we can help

The debate on the use of copyrighted works in AI training is just beginning. This ruling shows that the key will be to anticipate with a solid and well-documented legal strategy.

At ECIJA Chile we can assist you in designing and implementing legal strategies that allow you to take advantage of AI opportunities without exposing yourself to regulatory risks or litigation.