Russian Law on Artificial Intelligence: What Developers, Rightsholders, and Users Need to Know

From September 1, 2026, the law "On Support for the Development of Artificial Intelligence Technologies in the Russian Federation" (hereinafter referred to as the Law on AI) will come into force (with the exception of certain provisions). In this article, we will talk about how the new law will affect the developers of large fundamental models of artificial intelligence, their users, and copyright holders of copyright (related) rights.

What does the new law regulate?

Let's start with the fact that the Law on AI legalizes the conceptual apparatus, in particular, defines the concept of artificial intelligence as a set of technological solutions that allow you to imitate human cognitive functions and obtain results comparable to or superior to the results of human intellectual activity when performing specific tasks.

The ability of AI to achieve desired results is due to the fact that it can independently improve its functions, increase the accuracy of decisions by analyzing new data, and perform search for solutions even without a predetermined algorithm.

However, the Law on AI establishes the rules for the development, implementation, and use of not all artificial intelligence tools, but only giant neural networks, which belong to large foundational models of artificial intelligence.

Such models are used to perform a large number of different tasks, since they are trained and self-learned on the basis of given algorithms and data arrays and contain at least 1 billion parameters. Their functionality allows you to recognize and derive patterns, find the necessary information, make decisions, predict results for goals defined by a person, etc.

Thus, depending on the direction and specific functionality, the models allows the user to conduct a dialogue, generate texts and images, analyze documents and data, predict, program, as well as perform other visual, linguistic and search tasks.

Examples of products based on the use of foundational models are, in particular, Russian solutions - YandexGPT (developer - Yandex), GigaChat (developer - Sberbank), and foreign ones - Gemini (developer - Google, USA), ChatGPT (developer - OpenAI, USA), DeepSeek (developer - DeepSeek AI, China).

It is important to note that most of the provisions of the Law on AI contain general rules for the development, implementation, and application of foundational models, and a number of specific conditions will be established by executive authorities, in particular, the Government of the Russian Federation.

Next, let's analyze which provisions of the Law on AI directly affect the participants in the relevant legal relations.

Types of domestic foundational models and requirements for them

As a general rule, the Law on AI does not prohibit the use of foreign foundational models, but limits the scope of their application, and also establishes a number of advantages for domestic (national and sovereign) models. The status of a national or sovereign model is assigned by the Government of the Russian Federation if the foundational model meets the following general requirements:

• the developer is a Russian legal entity;
• data processing centers for responding to requests from models' users and storing data are located in Russia and belong to Russian legal entities;
• the model complies with domestic legislation and traditional spiritual and moral values of Russia. The procedure for such compliance will be established by the Government of the Russian Federation.

In addition, the model must meet special requirements depending on the status. A foundational model can obtain sovereign status if:

• the characteristics of the model at all stages of the life cycle are determined and changed by the Russian legal entity that develops the model;
• the developer of the model provided the opportunity to fully (technically and technologically) reproduce the entire development cycle of the model, including its training.

The model can obtain the status of a national if:

• the essential characteristics of the model (structure, software, configurable parameters) are determined and modified by the Russian legal entity – the developer of the model. At the same time, the list of necessary essential characteristics will be established by the Government of the Russian Federation;
• components that are used in the development of the model are distributed under an open license.

Benefits for national and sovereign foundational models

The Law on AI declares that to ensure the development of national and sovereign models, their developers are provided with appropriate measures of state support, including those aimed at stimulating the development, implementation and application of models (Part 1 of Article 6, Part 1 of Article 7 of the Law on AI). Presumably, these can be grants, tax benefits, and other financial advantages.

In addition, the developers of such models have the right to:

• participate in the activities of relevant industry organizations, the development of ethical rules for the use of AI technologies and the preparation of proposals for improving state policy in the relevant area;
• to train models on special conditions;
• carry out international cooperation, including for the implementation of joint research projects and the export of Russian AI technologies;
• insure property interests associated with the development, implementation and application of the models, as well as insure the risk of liability for harm caused by their use.

In addition, the Government of the Russian Federation may establish cases in which only national and/or sovereign models will be used (Article 5(1)(3) of the Law on AI). Thus, the developers of domestic models will also have an advantage in applying the model in certain areas of activity, for example, in banking and other areas of the financial market.

This rule will come into force on March 1, 2027. At the same time, until September 1, 2023, a transition period will be established for information systems in which the model is used, if they are created and (or) operated as of March 1, 2027.

Training of models on objects of copyright (related) rights

Separately, we would like to note an important advantage established for national and sovereign models – the right to train a neural network on objects of copyright and related rights. Thus, according to Part 3 of Article 10 of the Law on AI, domestic models will be able to use objects of copyright (related rights) for appropriate practical application, including their computer processing, extraction, comparison, classification, analysis of patterns, short-term recording in computer memory, etc.

The main condition is that the developer of the model has lawfully received a copy of such a work, or this object of copyright (related rights) is in the public domain and is not technically limited for analysis.

Thus, the Law on AI resolves a long-standing dispute between copyright holders and developers of neural networks on the free use of such objects as input data of neural networks for their training.

National and sovereign models will be able to freely train a neural network under the specified conditions from March 1, 2027, so the copyright holders of copyright and related rights objects who are against neural networks being trained at their facilities still have time to provide technical protection for their objects.

Who owns the rights to AI-generated content?

According to Articles 1228 and 1257 of the Civil Code of the Russian Federation, the author of the result of intellectual activity can only be a person whose creative work has created such a result.

At the same time, it follows from paragraph 1 of Article 1228 that technical, consulting, organizational, or material assistance, or assistance in the absence of creative contribution, does not give grounds for the emergence of copyright to the result of intellectual activity.

Thus, it is logical that a neural network (its algorithms) cannot be the author of the content and, accordingly, have copyrights to it.

In this regard, during the discussion of the draft Law on AI, the issue of assigning copyright to a person - a user who creates content using neural networks - was actively considered. It is assumed that in this case the user's creative contribution consists (cumulatively) of the idea of creating certain content, forming a prompt corresponding to the idea, selecting and editing the results, etc.

However, at the moment, neither the Civil Code of the Russian Federation nor the Law on AI answers the question of who owns the rights to the content generated using artificial intelligence (neural networks).

Instead, the new law introduces a rule that the person providing the opportunity to use models is obliged to notify the user of the ownership of the rights to the content obtained using this model, and how it can be used, accessed, uploaded, etc. (Part 1 of Article 10 of the Law on AI).

Given that such a notice affects, among other things, essential conditions – the rights to the result of intellectual activity and the methods of its use, it is logical to assume that this notice will be "embedded" as conditions in the user agreement.

Thus, the standard practice of not reading user agreements (according to statistics, 95% of Russians accept the terms of such agreements without reading) becomes vicious and can lead users of models to negative results, for example, to claims for compensation for illegal use of content.

The rule on notifying users will come into force on March 1, 2027.

Do you need to label the generated content?

At the stage of developing the draft Law on AI, the issue of the need to label content generated by the user using neural networks was discussed. However, in the final version of the new law, this issue is resolved differently.

For example, the Law on AI does not oblige users to mark the generated content, i.e., to notify that the result was created using a neural network. Instead, Part 1 of Article 9 of the Law on AI obliges models' developers, whose users create visual and/or audio content, to provide content creators with the opportunity to post an appropriate information warning. The format, content, and procedure for posting such a warning are established in the user agreement.

Also, the technical ability to post an information warning about generative content should be provided on social networks and other similar platforms (blog platforms, etc.), where users create personal pages and post information content in Russian and other languages of the Russian Federation. 

Such sites are required to provide users with the opportunity to label the generated content if the site can distribute advertising intended for Russian consumers, and the daily audience of users located in Russia exceeds 500 thousand people (Part 3 of Article 9 of the Law on AI).

It should be noted that Article 9 of the Law on AI formulates requirements for labeling content only in terms of the obligations of the neural network, but not users. It follows from the literal content of the norm that labeling is voluntary for users - content creators.

At the same time, experts recommend that before publishing generative content, you carefully read the terms of the user agreement, fearing that voluntary labeling may be transformed into mandatory by virtue of freedom of contract.

In conclusion, we note that the rules for labeling generative content will come into force on March 1, 2027.