Protecting intellectual property rights in generative AI training

Generative AI allows new content (text, computer code, images, audio and video) to be created in response to a user's request. The training of such AI often takes place on open data, which may contain IP-protected objects. This runs the risk of creating works that are like the copyrighted works of other authors.

Therefore, in May 2024, China and the EU adopted the first IP enforcement rules for generative AI training, which was prompted by a number of court cases where rights holders filed complaints about AI training based on their works without their authorization.

The experience of China

In May 2024, China released a draft “Basic Requirements for Generative AI¹ Security”. Generative AI service providers should:

1) Designate a person responsible for the observance of IP rights when using IP by the system and in the generated content.

2) Have an IP enforcement strategy, including a list of risks of IP infringement.

3) In case of IP infringement not to use infringing datasets for AI training, conduct infringement verification.

4) Provide a mechanism for users to complain to the vendor about AI rights infringements by generative AI.

Regarding the need to obtain permission from copyright holders for the content used for AI training, in February 2024, a court in Guangzhou held the owner of a website that provided a content generation service for money liable. The court found that the system operated in such a way that, at the user's request, an image was created that was confusingly similar to the plaintiff's intellectual property. The court's practice shows that AI that allows a user to make a request to generate content that involves copying IP or its individual elements will be treated as an infringement of IP law. In this case, the general rule on the prohibit on commercial use of IP without a² license from the rights holder applies.

At the same time, a court in China recognizes an AI user's copyright on AI- generated works. In November 2023, a Beijing court ruled on a copyright infringement case involving an image created using Stable Diffusion. The AI-generated images are “works” under copyright law because they belong to the field of literature, art, are original and represent the result of human intellectual activity. The author is recognized as the plaintiff, who entered a query of keywords (type of image, depicted object, environment and style) and further adjusted the query. The court ruled out the authorship of the developers of the AI system on the generated image.

The experience of EU, France,

and Italy

Ita Adopted in May 2024, the EU AI Act establishes the obligation of generative AI providers to:

1) Implement compliance with copyright law (e.g., check the data used by AI for protected IP objects).

2) Publish a detailed description (summary) of the content used to train the AI model (part 1 of Article 53), e.g. listing the main datasets, including an indication of major private or publicly available databases or archives. It is planned to develop a form for such a summary.

In the EU, there is a right for research organizations and cultural heritage institutions to the IP use without the permission of the right holders for scientific purposes. Similar to Article³ 23 of the Copyright Law of China.

Among the EU countries, in September 2023, France was the first to introduce Bill No.⁴ 1630 on mandatory obtaining the consent of right holders when using IP objects for AI systems, including training of generative AI systems. Without author’s authorization, it is possible to use only for non-commercial purposes. This reduces the risks of unlawful commercial use of IP objects while preserving the possibility of free use of such objects for socially useful purposes.

If a work is generated by an AI “without direct human intervention”, the right holders are the authors or right holders of the works that made the generation of the work possible (Art. 2). However, it remains unclear how the specific works that became the basis for the generation should be defined, how should the copyright on the work generated by the AI be assigned?

The French bill proposes to introduce collective management of AI-generated works through collective management organizations (to collect royalties). The draft law does not solve the issue of using for training works whose authors could not be identified. Such a proposal carries the risks of misidentifying the authors of the work, infringing the authors' rights to the name and to remuneration for the use of their works.

France also proposes to make it compulsory to mark that a work has been generated by an AI system - a similar rule has been introduced in Italy's draft AI law (April 2024). A sign or marking with the abbreviation⁵ “IA” (intelligenza artificiale) must be used. Regarding the use of protected IP for AI training, Italy, unlike the EU and France, has offered a different approach - the right to freely use legally accessible data for training AI systems, unless such use is expressly prohibited by the right holders.

The US experience

No specific regulation has been enacted in the US. As a rule, the doctrine of “fair use” of copyrighted data applies - it is possible to use IP objects without a license in the circumstances defined:

1) Purpose (e.g., educational).

2) The nature of the use (noncommercial). This does not mean that every non-commercial use is recognized as bona fide and every commercial use as non-bona fide. For example, a “transformative use” (adding something new) would probably be considered fair use. Using a combination of image and text to train an AI is essentially “transformative” over the original data, which is theoretically recognized as fair use.

3) The character of the data itself: the greater the creative element, the lower the chances of fair use being recognized.

4) The amount of source data used.

5) The impact on the market and value of the protected data - whether unlicensed use harms the existing market (e.g. by displacing sales of the original) or the future market (the original will become less popular).

In 2023, lawsuits were filed in the US by US writers against OpenAI for using their texts for ChatGPT training and copying works without permission; a lawsuit by artists against Stability AI, Deviantart and Midjourney for unauthorized use of copyrighted images for AI training to produce more works of the same type without the consent of the original image authors. The complaints have been accepted for review, but rulings have not yet been given.

Russia’s experience

There is no regulation of IP for AI training in Russia.

However, it is possible to apply Article 1274 of the Civil Code of the Russian Federation - the right to use the publicized work or its part without the consent of the right holder and without payment of remuneration for informational, educational or cultural purposes. It is required to specify the name of the author and the source. That is, if the above conditions are met, the use of IP objects may be qualified as quoting, and IP objects may be used for training generative AI for commercial purposes, which entails infringement of IP right holders.

Thus, it is possible to highlight 3 approaches of countries to the use of data for AI training:

1) Requirement for AI service providers to comply to protect IP and publish information on data used for AI training (EU, France and China). In addition, generally, it is required to obtain authorization from right holders for commercial use of IP, without authorization - only for non-commercial purposes.

Courts in China and a bill in France propose to extend a general rule on obtaining authorization from right holders for commercial use of IP for AI training.

2) No special requirements for AI operators (US) while allowing bona fide use of IP for AI training without a license.

3) Free use of AI without the right holders' authorization, except for an explicit prohibit on use by the right holder (Italy).

The Russian approach to the right to use works for training AI without authorization is similar to the EU and French approaches. Russia has no mechanisms to protect right holders from unauthorized use of IP objects for training generative AI systems.

To create such mechanisms, it is possible to supplement Part 4 of the Civil Code of the Russian Federation with norms:

− The right of the author/right holder to forbid the use of the work for training AI systems / algorithms. − Marking of works created using

AI/algorithms.

From the monitoring issue No. 5, May 2024. Download the full issue (PDF) · issue page at the Gaidar Institute