Russian Bill on Supporting Development of Artificial Intelligence Technologies

On July 7, 2026, the State Duma passed in the first reading Draft Federal Law No. 1271570-8“On supporting development of artificial intelligence technologies in the Russian Federation” (hereinafter the “Bill”), submitted by the Government of the Russian Federation on June 25, 2026.

This is a substantially revised version of the bill “On fundamentals of state regulation of areas of application of artificial intelligence technologies”, published by the Russian Ministry of Digital Development in March of 2026 (hereinafter the “March Edition”). On July 8, 2026, the Bill was adopted in the State Duma in the second and third readings and was forwarded to the Federation Council.

In this review, we analyze the key provisions of the Bill, how it is different from the March Edition and highlight the practical implications for anyone involved in AI in any capacity.

WHAT DOES THE BILL REGULATE?

The Bill introduces basic concepts in the field of AI into Russian legislation for the first time (Article 3 of the Bill): “artificial intelligence,” “large fundamental artificial intelligence model” (hereinafter “LFM”), “data composition,” “model training,” “developer,” “Russian legal entity” and others.

At the same time, the scope of regulation has been significantly narrowed compared to the March Edition (Article 1 of the Bill): The Bill applies only to LFMs – computer programs designed to perform a wide range of intellectual tasks at a level comparable to or exceeding the results of human intellectual activity and containing at least 1 billion parameters. The explanatory note cites YandexGPT, GigaChat, ChatGPT, Gemini and DeepSeek as examples of LFMs.

In other words, the Bill does not apply to small and highly specialized models (such as computer vision systems) or to application services built based on third-party LFMs.

WHAT HAS CHANGED COMPARED TO THE MARCH EDITION?

The bill has been reduced from 21 to 13 articles and has shifted its focus from government regulation of the industry to its support. The most controversial provisions of the March Edition have been removed from the text:

• The category of “trusted” AI models and the registry of such models approved for use in government information systems and at significant critical information infrastructure (CII) facilities;
• Obligations of AI system operators, AI service providers and users, including specific liability provisions that were intended to introduce the standards of “knew or should have known” and “exhaustive measures,” the vagueness of which would have created significant legal risks for market participants;
• Mandatory labeling of AI-generated content and the obligation of platforms to verify its presence;
• Restrictions on “cross-border” AI technologies: foreign neural networks are neither prohibited nor restricted;
• Requirement that sovereign and national models be trained exclusively on data of Russian origin;
• Special regime and support measures for data centers and computing infrastructure for AI.

Some excluded issues (such as labeling, liability and certain copyright aspects) have not been definitively resolved, but have been deferred to future subordinate regulations and sector-specific legislation.

SOVEREIGN AND NATIONAL MODELS

The Bill proposes dividing LFMs into “sovereign” and “national” models. Both categories must comply with general requirements (Parts 2 and 3 of Article 6 of the Bill):

• The developer is a Russian legal entity (controlled by the Russian Federation, a constituent entity of the Russian Federation, a municipal entity and/or Russian citizens who do not hold foreign citizenship and/or persons they control jointly or individually);
• Responses to user requests and data storage must be carried out in data centers located within the territory of the Russian Federation and owned by Russian legal entities;
• The model has been certified as compliant with Russian legislation and traditional Russian spiritual and moral values in accordance with the procedure to be established by the Russian Government.

The sovereign model further requires that the development and modification of the model’s characteristics at all stages of its life cycle be carried out by a Russian legal entity and that full technical and technological reproducibility of the development cycle be ensured, including training (mention of “matrix of initial coefficients” was deleted in the second reading). Essentially, these are models trained “from scratch.”

The national model is a less restrictive category: the use of components created outside the Russian Federation, including third-party machine learning models, is permitted, provided they are distributed under an open license (for example, fine-tuned versions of open-source models).

The Russian government will have the authority to specify cases in which only sovereign and/or national models may be used (in the banking sector and other areas of the financial market – subject to approval of the Bank of Russia) (Clause 3 Part 2 Article 5 of the Bill). This is expected to affect primarily the public sector, government information systems and critical information infrastructure (CII) facilities; the Bill does not require the private sector to switch to domestic models.

Developers of sovereign and national models will have access to government support measures and to data from government information systems for training the models (Article 7, Clause 5 Part 2 and Part 5 of Article 5 of the Bill). At the same time, they are required to take measures to ensure the security of the model, establish rules for its operation and maintain technical documentation (Article 8 of the Bill). It should be noted, however, that the procedure for confirming a model’s compliance with traditional Russian spiritual and moral values has not yet been established; therefore, it is currently difficult to assess the effort required to obtain the status of a sovereign or national model: the actual appeal of these categories will largely depend on the content of future resolutions of the Russian Government.

LABELING AI-GENERATED CONTENT

Instead of mandatory labeling, the Bill establishes a voluntary model (Article 9 of the Bill): a person who creates audio and/or visual materials using AI-powered systems must be given the opportunity to post an informational notice regarding the use of AI technologies.

The format, content and procedure for posting the notice are to be determined by an agreement between the user and the party providing access to the model. In the second reading, the circle of obligated persons was clarified: the obligation to provide users the possibility of such labeling is imposed on owners of web sites, information systems and programs that users use to disseminate information through personal pages on which advertising can be distributed and access to which exceeds 500,000 daily users from Russia – in essence, to owners of social networks according to the criteria of Federal Law No.149-FZ dated July 27, 2006 "On information, information technologies and information protection" (Part 3 Article 9 of the Bill). The obligation is formulated imperatively ("obligated to provide") and covers any information created using LFMs, not just materials in audio and/or visual form. In our view, this approach is a compromise: it reduces the regulatory burden on the industry, but effectively leaves the issue of informing the audience to the discretion of users and makes it subject to the terms of their agreements with the services.

Please note that mandatory labeling of certain types of content may be introduced later through subordinate legislation or industry regulations; therefore, it is advisable for generative AI services to incorporate the technical capability for labeling into their product architecture in advance. We should note that the amendment on mandatory marking under a procedure established by the Russian Government was rejected in the second reading.

INTELLECTUAL PROPERTY

Article 10 of the Bill contains two key provisions.

First, entities providing access to LFMs are required to notify users of the ownership of results of intellectual activity (RIA) obtained using the model, as well as the terms and conditions for accessing, using and saving (if technically possible) such results (Part 1 Article 10 of the Bill). The procedure for such notification is not specified; we believe that, in practice, these issues will be addressed in user agreements, which creates a risk for users that the service provider may impose terms under which it retains all rights to the generated content. We also note that due to the principle of human authorship (Article 1228 of the Russian Civil Code), the model itself cannot be recognized as the author of the generated results: AI remains an object, rather than a subject of civil rights.

Second, an exception to the exclusive right is introduced (analogous to the European TDM exceptions^[1]): accessing protected works, including their extraction, comparison, classification and analysis, as well as short-term storage in computer memory, exclusively for the purpose of training a sovereign or national LFM, is not deemed a violation of copyright, provided that one of the following conditions is complied with:

1. The developer lawfully obtained a copy of the work; or
2. The work has been made available to the public and/or is accessible for analysis without technical restrictions (Part 2 Article 10 of the Bill).

Compared to the March Edition, this exception to the exclusive right applies only to objects of copyright and related rights (patent rights are not mentioned) and only to the training of sovereign and national models.

In our view, this provision remains the most controversial: the exception to the exclusive right is introduced by a special law without making appropriate amendments to Part Four of the Russian Civil Code, which establishes an exhaustive list of cases in which works can be used freely. The risk of such a conflict also existed in the March Edition. Questions also remain unresolved regarding remuneration for rightholders, the mechanism for their opt-out from the use of lawfully obtained copies, as well as the provision’s temporal application – specifically with regard to models that have already been trained. We believe that without resolving these issues, the provision will create grounds for disputes between rightholders and developers, and the regulatory gaps will need to be compensated by courts.

We should note that the issues of regulating relations in the field of intellectual property rights, as well as cases of free use of works, are established by the Civil Code and international treaties to which the Russian Federation is a party. In this regard, it cannot be ruled out that the provisions of the Bill restricting exclusive rights to works may be regarded as contrary to the norms of legislative acts having greater legal force.

LIABILITY

Specific provisions on liability have been excluded from the Bill: the parties to the relationship are liable in accordance with the legislation of the Russian Federation (Article 11 of the Bill). Thus, the issue of allocating liability for harm caused by the results of the models’ operation (including so-called “hallucinations”) between the developer, the entity that implemented the AI and the user remains unregulated and will be resolved in accordance with the general provisions of civil, administrative and criminal law.

HOW HAS THE BILL CHANGED IN THE COURSE OF EXAMINATION IN THE STATE DUMA?

On July 3, 2026, the State Duma Committee on Information Policy, Information Technology and Communications (hereinafter the “Committee”) recommended adopting the Bill in the first reading, subject to revisions to address the comments raised.

While supporting the concept of the Bill as a whole, the Committee noted that its provisions need to be brought into line with the provisions of existing federal laws, as well as refined in terms of legal and technical aspects. In particular, the Committee pointed out the following:

• Title and subject matter – the title of the Bill refers to supporting the development of AI technologies in general, whereas its subject matter is limited to the regulation of LFMs (Part 1 Article 1 of the Bill); therefore, the title should be aligned with the subject matter of the regulation;
• Terminology – the definition of the term “artificial intelligence” must be brought into line with the definition already established in Federal Law No. 123-FZ of April 24, 2020, and the legal meaning of the term “parameter” must be clarified either in the Bill or at the subordinate legislative level. In addition, the text uses terms that are not defined in the legislation (for example, “technological reproducibility,” “key economic sectors”);
• Model-based decision-making – the reference to an LFM making decisions based on human-defined objectives (Clause 2 Article 3 of the Bill) should be deleted, as this may involve the performance of legally significant actions;
• Training datasets – the Bill should be amended to stipulate that the creation of training data sets must comply with Russian legislation, including copyright law;
• Model localization – the Committee notes the absence of a requirement to host sovereign models in Russian data centers: currently, the Bill provides only for the storage of responses to user requests and related data;
• Areas subject to special regulation – the list of areas for which specific rules for the use of LFMs may be established (Part 2 Article 1 of the Bill), should be expanded to include preliminary investigations, the administration of justice and the enforcement of criminal penalties, and the obligation of model developers and users to preserve state and official secrets should be enshrined.

On July 8, 2026, the Bill was adopted in the second and immediately in the third reading: the amendments recommended for adoption by the Committee after consideration of the Bill in the first reading were approved, and those recommended for rejection were rejected.

At the same time, the Committee's comments were only partially taken into account. By the second reading, the terminology was clarified: for example, the concept of "data set" was replaced by the concept of "data composition." The definition of artificial intelligence was supplemented with an indication of the ability to independently improve its functions and improve the accuracy of decisions through the analysis of new data (by analogy with Federal Law No.123-FZ of April 24, 2020). Individuals were excluded from the definition of developer (only an individual entrepreneur or a legal entity can act as a developer). The definition of a "Russian legal entity" has been moved from Article 3 to Part 4 Article 6 of the Bill. The concept of "state support measures" was eliminated.

At the same time, the name of the Bill was not changed, the reference to model decision-making remained, there was no requirement to place the sovereign models themselves in Russian data centers, the list of special regulation areas was not supplemented (Part 3 Article 1) and a special rule on the formation of training data sets/compositions while complying with copyright legislation was not included. In the third reading, only legal and technical amendments were made.

WHEN WILL IT TAKE EFFECT?

The Bill provides that the new law will take effect on September 1, 2026 (Part 1 Article 13 of the Bill), and its key provisions (cases of exceptional use of sovereign and national models, access to government data, model criteria, developers’ obligations, labeling and intellectual property rules) will take effect on March 1, 2027 (Part 2 Article 13 of the Bill).

A transition period is also established (Part 3 Article 13 of the Bill) – until September 1, 2032, the requirements for the exclusive use of sovereign and/or national models will not apply to information systems created and/or operated at the time the relevant provisions take effect, provided that data is processed and stored within the territory of the Russian Federation.

By November 1, 2026, the Russian Government must adopt five resolutions (the agency in charge is the Russian Ministry of Digital Development, Communications and Mass Media, with the participation of the FSB, the Federal Service for Technical and Export Control, the Ministry of Culture and Rosfinmonitoring):

• On cases of exclusive use of sovereign and national models;
• On the procedure for accessing government data;
• On rules for confirming compliance with traditional values;
• On essential characteristics of the national model; and
• On the procedure for registering and assigning statuses.

Simultaneously with the adoption of the law, the State Duma recommended that the Russian Government provide information on the regulatory legal acts adopted in compliance with the law by March 1, 2027. It is also possible to establish the specifics of the application of Articles 3 and 6 of the Bill within the framework of experimental legal regimes. In the second reading, it was clarified that the programs of such regimes are to be coordinated with the Russian Ministry of Finance (Part 4 Article 13 of the Bill).

WHAT SHOULD BUSINESSES DO?

For most companies using AI services (including foreign ones), the Bill does not impose any new obligations. Nevertheless, we recommend:

• That companies working with the public sector, CII and in the financial sector conduct an inventory of their implemented AI solutions, ensure that data is processed and stored within the Russian Federation (a condition for the transition period) and plan in advance for a possible transition to sovereign or national models;
• Developers of LFMs seeking sovereign or national model status should monitor the adoption of Russian Government resolutions, prepare technical documentation and operating rules and assess the reproducibility of the development cycle and the licensing compliance of the components used in the AI model;
• Generative AI services should establish the technical capability to label content and update their user agreements regarding notifications of rights to generated digital content and the terms of its use;
• Owners of web sites, information systems and programs with a daily audience over 500,000 users in Russia should provide users with the technical ability to post notices about the use of LFMs when distributing content created with their help;
• Rightholders should consider implementing technical measures to restrict access to content (e.g. machine-readable bans): the absence of such measures will allow developers to train models on such content without the rightholder’s consent and without paying compensation. After the second reading, the significance of such measures has grown: if there are technical restrictions, this unequivocally rules out the free use of an object, even if it is publicly available.

We will continue to monitor the examination of the Bill in the Federation Council and will keep you informed of further developments. The CLS intellectual property group team is ready to assist you in assessing the impact of the new regulations on your business.

www.cls.ru @LegalRecipes LegalRecipes в MAX

• ^[1] A TDM exception is an exception for the purposes of text and data mining provided for in Articles 3 and 4 of the EU Directive 2019/790 on copyright and related rights in the digital single market. It allows the reproduction and extraction of legally accessible works and other protected objects for the purposes of automated analysis of text and data in digital form, aimed at obtaining information, among others, about patterns, trends and correlations. The exception applies unconditionally to scientific research (Article 3 of the Directive). In other cases, the copyright holder has the right to prohibit such use, including by using machine-readable means (Article 4 of the Directive). Source: https://eur-lex.europa.eu/eli/dir/2019/790/oj