It should be noted that the original version of the draft law (the draft law “On the Principles of State Regula-tion of Areas of Use of Artificial Intelligence Technology in Russia”) was criticised by the Presidential Coun-cil for the Codification and Enhancement of Civil Legislation. Experts pointed out, in particular, that certain provisions of the document contradict the Civil Code and undermine the foundations of copyright.
As a result, the draft no longer includes a ban on cross-border AI solutions, the concept of “trusted mod-els”, provisions on liability in connection with the use of AI, or provisions governing copyrights to works created by AI (in particular, the drafters abandoned originality as a criterion for copyright protection). The focus has shifted towards supporting domestic developments, creating preferential regimes for training models and introducing special statuses (sovereign and national models), which will grant access to the public sector and regulated industries.
Application areas of AI
The Draft Law does not regulate all AI solutions. Its scope is limited exclusively to big foundation models (“BFMs”), specifically those containing at least 1 billion parameters (today, BFMs with a much larger num-ber of parameters — from 3 billion and above — are used in business).This means that highly specialised AI solutions, traditional machine learning, computer vision and applied algorithms fall outside the scope of regulation. If a business uses or develops, for example, narrow ML models for specific purposes (such as recommendation systems for websites), then this is not covered by the Draft Law.
Newly introduced sovereign and national model statuses
The Draft Law establishes two categories of BFMs: the sovereign model and the national model. Such sta-tus entitles the holder to government support, as well as simplified access to datasets.A sovereign BFM
The sovereign model implies that the developer must be a Russian legal entity. For the purposes of the Draft Law, a Russian legal entity is defined as a legal entity controlled by the Russian Federation, by a constituent territory of the Russian Federation, by a municipal authority, by a citizen of the Russian Federa-tion who does not hold the nationality of any other state, or by persons controlled by any of them. Control is defined as the ability to directly or indirectly exercise more than 50% of the voting rights attached to the voting shares or participating interest in the charter capital. This wording opens up opportunities for estab-lishing joint ventures with foreign partners in the field of BFM development, which may subsequently be granted sovereign or national status.
However, the development of the model, including the definition and modification of its characteristics at all stages of its life cycle, must be carried out by a Russian legal entity which is the developer of that model. The Draft Law also requires the development cycle to be fully reproducible from a technical and technologi-cal perspective, including training and the matrix of initial coefficients.
A national model
The national model must also be developed by a Russian developer. However, as opposed to the sovereign model, the Draft Law permits the use of components made in the Russian Federation or outside it, including other BFMs, in the development of the national model, provided that such components are distributed on the terms of an open-source licence. This is important because businesses now frequently use open-source BFMs and build their solutions around them (for example, Qwen, DeepSeek, etc.).
For both categories of models, there is a requirement that user requests must be processed and data must be kept in data centres located within the Russian Federation and owned by Russian legal entities.
In addition, both the sovereign and national models must be verified for compliance with Russian law and traditional Russian moral values. The Government of the Russian Federation has been tasked with estab-lishing the verification procedure, as well as the procedure for the registration and assignment of the status of sovereign and national models.
Intellectual property
The Draft Law contains specific provisions on intellectual property in the context of the use of BFMs.Those who make BFMs available for use are obliged to inform the user of the ownership of the rights to the results of intellectual activity obtained using such a model. Users must also be informed of the terms and conditions governing the access to, use of and downloading of such results, provided that it is technically possible to download such results.
The Draft Law also contains a provision of critical importance to BFM developers. Reproduction by means of short-term storage in computer memory is permitted solely for the purpose of training sovereign and/or national BFMs, provided that the following conditions are met: (a) the developer must use a copy of the work obtained lawfully, or (b) the work protected by copyright or related rights must have been made public-ly available and/or be accessible for analysis without technical restrictions (in other words, circumventing paywalls, subscriptions and other content protection measures is prohibited). What is more, the Draft Law does not regulate the issue of remuneration, and it is assumed that such use will be free of charge.
Therefore, in essence, this means that free training on open data is introduced if a developer creates a sovereign or national BFM. However, the Draft Law does not regulate matters relating to personal data that may be contained in such open sources, which means that the provisions of the Federal Law “On Personal Data” continue to apply in full to such open data (including the obligation on the developer of the BFM to ensure an appropriate legal basis).
Labelling AI-generated content: from an obligation to an option
The Draft Law specifically regulates the labelling of information materials produced using BFMs. This refers to informational material in audio and/or visual form, which is important for businesses that actively use the creation of marketing content in their operations.A person using BFM to create such material must be given the opportunity to include a warning stating that AI technology has been used. The format, content and placement of such a warning must be determined by agreement between the person using the model and the person making it available for use.
Social media platform owners must also ensure that users who share audio and/or visual content created using BFMs are able to include a warning about the use of artificial intelligence technology.
In other words, when creating content for marketing purposes, it is not necessary to state that such content was created using AI. This may be determined by the rules of a particular platform (social media, video sharing platforms, etc.), but it is a matter of private law rather than a public obligation on the part of busi-nesses.
Obligations of the developers of sovereign and national BFMs
The developer is obliged to take organisational and technical measures to ensure the security of the BFM. The Draft Law does not set out a specific list of such measures; therefore, it is likely that they will be clari-fied in subsequent regulations or determined with reference to the nature of the model and its application area.The developer must also define the operating rules for the BFM. Such rules should set out the restrictions and conditions for using the model, as well as the procedures for updating and decommissioning it.
In addition, the developer is required to maintain technical documentation describing the key parameters and limitations of the BFM, which must be sufficiently comprehensive to enable an assessment of the in-formation security of the BFM’s application.
