The State Duma adopted in the first reading a framework bill "On supporting the development of artificial intelligence technologies in the Russian Federation," which for the first time establishes basic concepts in the field of AI and introduces categories of sovereign and national large models. These include systems such as YandexGPT, GigaChat, ChatGPT, DeepSeek, and Gemini. The government will have the right to determine areas where only Russian models can be used, establish state support measures, and regulate developers' access to state data. More details about the bill can be found in the material from "Pervy Tekhnichesky".
The law will not affect every AI program
The bill regulates not all algorithms and neural networks, but large fundamental artificial intelligence models, or LFM AI. These are programs capable of performing a large number of different intellectual tasks, providing information, making decisions, or predicting results based on human-defined goals.
To fall under this definition, a model must contain at least 1 billion parameters and be used as a basis for creating and refining other software products.
In the explanatory note, such systems are described as programs trained on gigantic volumes of diverse data — texts, images, sounds, and other information.
As examples of solutions based on large fundamental models, the document's developers cite Russian YandexGPT and GigaChat, as well as foreign products from Google, OpenAI, and DeepSeek. The law will also apply to providing access to such systems to users located in Russia.
At the same time, the legislation proposes to establish a general definition of artificial intelligence. It is understood as a complex of technological solutions that imitates human cognitive functions, including self-learning and finding solutions without a predefined algorithm. The results of such a system's work must be comparable to or exceed human intellectual activity.
Sovereign and national models will appear
One of the main innovations of the bill is the division of large models into sovereign and national. Both categories will be able to claim state support, but the requirements for them will differ.
The developer of a sovereign model must be a Russian legal entity. It is obliged to independently determine and change the characteristics of the system at all stages of its life cycle. In addition, the developer must ensure full technical and technological reproducibility of the entire model creation cycle, including its training and the matrix of initial coefficients.
The responses of a sovereign model to user requests must be formed in data processing centers located in Russia and owned by Russian legal entities. Data storage will also be required within the country.
Additionally, the model must pass a check for compliance with Russian legislation and traditional Russian spiritual and moral values. The procedure for such a check will be determined by the government.
The requirements for a national model are softer. Its developer must also be a Russian legal entity that controls the essential characteristics of the system: structure, software, and configurable parameters.
However, when creating a national model, it is allowed to use Russian and foreign components, including other large models, if they are distributed under an open license. The preparation of responses and storage of information, as in the case of a sovereign system, must take place in Russian data centers owned by Russian legal entities.
A national model will also require confirmation of compliance with legislation and established values. The government will separately establish rules for accounting for such systems and assigning them official status.
Where only domestic systems can be left
The government will have the right to establish cases where only sovereign or national large models are allowed to be used. For the banking sector and other segments of the financial market, such decisions will need to be coordinated with the Bank of Russia.
The Cabinet of Ministers will also determine possible exceptions to this requirement. There is no direct general ban on foreign neural networks in the bill: specific areas, organizations, and information systems must be named later in by-laws.
The explanatory note specifically mentions sensitive objects, including state information systems. For them, the government will be able to introduce mandatory use of sovereign or national solutions.
At the same time, the Cabinet of Ministers will have the right to establish industry requirements for preventing risks from the use of large models. Regulation should be built taking into account the probability and scale of possible harm to human life and health, property, business reputation, the environment, security, and technological independence of the state.
Developers may be given data from state systems
Training large models requires huge amounts of information. The bill creates a mechanism allowing developers of sovereign and national systems to access data sets from state sources.
A similar mechanism can be introduced by regions. The highest executive body of a Russian subject will have the right to determine the procedure for accessing data from regional authorities, state institutions, unitary enterprises, and organizations under the direct or indirect control of the region.
Access conditions, the composition of data sets, and the procedure for their provision must be established separately and agreed upon with security agencies.
What benefits are promised to neural network creators
The government will be able to introduce financial, property, guarantee, and information support measures. Specific amounts, benefits, and methods of obtaining them are not specified in the bill — they must be determined by by-laws.
Developers of sovereign and national models will have the right to participate in the work of specialized industry organizations, create mechanisms for ethical regulation, and prepare proposals for changing state policy in the field of artificial intelligence.
They will also be allowed to participate in joint international research, export domestic solutions, and insure property interests related to the development and application of models. Separately, the possibility of insuring liability for harm caused by the use of such systems is provided.
Developers will be responsible for security
Along with support, the bill introduces obligations for companies creating sovereign and national systems.
The developer must take measures to ensure the security of the model. They will need to establish rules of use, define restrictions and conditions for using the system, and describe the procedure for its updating and decommissioning.
In addition, the company is obliged to maintain technical documentation. It must specify the key parameters and limitations of the model to an extent sufficient to assess the information security of its application.
For violating the requirements of the law and regulatory acts adopted on its basis, market participants will be held liable in accordance with Russian legislation. Separate offenses and the amounts of possible fines are not named in the project itself.
AI content labeling will not automatically become mandatory
A separate article is devoted to images, videos, and audio materials created using large models. In its current version, the bill does not introduce an obligation to label all such content.
Instead, a person using the model to create audio or visual material should be given the opportunity to place a warning about the use of artificial intelligence technologies.
The format, content, and procedure for placing such a warning are proposed to be determined by agreement between the user and the company providing access to the model.
Social media owners will also have to provide users with the technical ability to add a corresponding warning to distributed images, audio recordings, and videos.
When the new rules will come into effect
The main part of the law is expected to come into force on September 1, 2026. Other significant requirements will take effect later — on March 1, 2027.
The entry into force of norms on the mandatory use of sovereign or national models in certain cases, industry requirements for risks, access to state data, assignment of special statuses to models, developer obligations, material labeling, and the use of intellectual property objects has been postponed to this date.
A long transitional period is provided for existing information systems. Until September 1, 2032, the requirement for the exclusive use of sovereign or national models will not apply to systems created or operated at the time the relevant norms came into force. The condition is that data processing and storage must be carried out on the territory of Russia.
In addition, experimental legal regimes with special rules for defining and classifying large models can be created for individual projects.
However, the main decisions are yet to be made by the government.
Read more on the topic:
- The Kremlin Discusses a Complete Transition to Sovereign AI: From Model Development to Implementation in Defense and the Economy
- Neural Networks to Be Licensed: Media Market Demands Payment for Books, Music, and Animation
- Neural networks have become a line in resumes: employers almost three times more often require AI skills