Protection of intellectual property rights in AI training

In 2025, the search for optimal regulation of intellectual property rights (hereinafter - IPR) in AI training continues, ensuring a balance between the protection of rights holders and the needs of developers. In Monitoring No. 5 of 2024, we have already considered approaches to addressing this issue in China, the EU, France, Italy, and the USA.

The US experience

In February 2025, a court decision was issued in the case of Thomson Reuters and West Publishing v. Ross Intelligence for copyright infringement of the legal annotations and classification system for law enforcement decisions (judicial, administrative, etc.) of the¹ Westlaw platform. Ross created an AI-based legal research tool by illegally using plaintiffs' materials to train its system.

The court found that Ross directly copied data from Westlaw in violation of the Copyright² Act, finding that Thomson Reuters owned the copyright not only in the texts, but also in their selection and organization. The court rejected³ Ross's fair use of content argument because the AI was used for commercial purposes and⁴ recycled the data in whole, not in part. Ross created a competing product based on illegally obtained content owned by Thomson Reuters and generated its own data from it in the same market.

The court also confirmed that Westlaw's AI-generated annotations had sufficient originality to be protected by copyright. That is, the owner of the AI system (Thomson Reuters) is recognized as the copyright holder of the generated content. Thus, the use of data in AI without the permission of the copyright holders violates the law, even with reference to fair use.

Also in February 2025, California published a bill to enhance transparency in the use of data for AI training. It is proposed to require generative AI developers to document the use of copyrighted materials in AI training, identify copyright owners, and keep data on the facts of such data use and its owners during the commercial use of the AI model, and 10 years after. Developers must provide a list of the materials used within 7 days upon the request of⁵ the copyright holder.

The UK experience

In February 2025, public hearings were held on plans to protect IP rights in the use of⁶ AI. The UK is considering four approaches to IP regulation in the context of AI training: (1) no specific regulation; (2) compulsory licensing; (3) the right for anyone to use the data for AI training; or (4) the same right, but with the right holder's option to prohibit use, and the developer's obligation to ensure transparency and accountability for the use of the data.

The UK has chosen the 4th approach accordingly, it is planned to introduce a model where authors will be able to prohibit the use of their works in advance, and in the absence of such a ban their content can be freely used. This will allow copyright holders to earn money on licensing, and developers - to work with large amounts of data. AI developers should disclose what data their models are trained on (similar to the norms of the EU AI Act).

Existing copyright protection, such as⁷ robot bans, is insufficient. They can block access to publicly accessible sites on the Internet, but do not protect individual works. Therefore, the UK is exploring ways to strengthen control over data through metadata⁸ with a ban on content collection, development of standards for recording access rights, copying, etc., and centralized registries of restrictions. It is proposed to transfer rights to collective management organizations for licensing and payment distribution.

Russia’s experience

Russia lacks legal regulation of IP rights protection in AI training. It was planned to approve an updated Concept of AI regulation in February 2025, but there is no information about such approval. The draft Concept sets the task of working out the issues of protection of AI rights to works created with the use of AI, labeling of such works for transparency of their origin. It is proposed that a person or organization that controls the system and makes decisions on its use should be responsible for⁹ the results of AI work. In other words, there is no special regulation in Russia now, however if the plans described above from the draft Concept are implemented, the Russian approach will include granting authorizations “by default” and with information disclosure requirements, as it is accepted in the EU and planned in the UK.

From the monitoring issue No. 2 (14), February 2025. Download the full issue (PDF) · issue page at the Gaidar Institute