According to the bill, a large foundation model is software capable of performing numerous intellectual tasks at a human-like proficiency. The definition includes systems that supply information, make decisions, or predict outcomes based on predefined objectives. To qualify, a model must contain at least 1 billion parameters. The framework emphasizes principles such as technological sovereignty, human rights, individual choice, security, and adherence to Russian legal standards.
The legislation establishes distinct categories for sovereign and national models. A sovereign model must originate from a Russian legal entity and stay under domestic control, utilizing data centers within Russia. Developers are required to maintain the ability to reproduce the entire development process, including training and original parameters. A national model adheres to similar ownership and data processing rules but may incorporate foreign components distributed under open licenses.
Recognition of Domestic Models as Official
The government can implement support measures for entities involved in creating, deploying, or managing qualifying foundation models. Such measures might include access to state-held data essential for training purposes. Additionally, authorities can designate specific environments where only sovereign or national models are permitted to operate, particularly in government information systems and sensitive sectors. Separate laws and presidential decrees could establish regulations concerning defense, security, public order, and property protection.
Websites, applications, and social media services with over 500,000 daily users will face a new requirement for content labeling. These platforms must offer users an option to mark AI-generated audio or visual content. The regulation provides a mechanism for users to append an informational notice to qualifying content, with developers and users agreeing on the label’s format through terms of service. Platforms are not mandated to automatically label all items but must provide the technical capability to do so.
Copyright and Regulatory Frameworks in Development
AI service providers are required to inform users about the ownership rights of generated material. They must also clarify access conditions and whether users can transfer or reuse the content. The legislation separately addresses the use of copyrighted works for machine learning, considering analysis activities like extraction, comparison, classification, and pattern recognition lawful when developers have obtained proper legal access. Training on protected works is permitted when no technical restrictions are bypassed.
Most provisions are set to come into effect on September 1, 2026, pending presidential approval and official publication. The rules concerning the status of domestic models, developer responsibilities, content labeling, and intellectual property will commence on March 1, 2027. Existing systems will have a transition period until September 1, 2032, during which their data processing and storage must stay within Russia. Until the bill is officially enacted, it remains an approved draft rather than a fully enacted law.