MOSCOW, RUSSIA / RankWire.AI / – The Federation Council sanctioned a comprehensive legislative framework for artificial intelligence on July 17, establishing guidelines for large foundational models within Russia. The legislation specifies the technologies it covers and grants authority to government agencies. It also sets standards related to model ownership, local data storage, user notifications, and content generated by AI. The bill was passed by the State Duma on July 8 and awaits presidential approval and official publication to become law at the federal level.
The legislation defines a large foundation model as software capable of executing various intellectual tasks at a level comparable to humans. Such a system must have at least 1 billion parameters. These models can provide information, make decisions, or predict outcomes based on human-set objectives. The framework emphasizes principles such as technological sovereignty, human rights, personal choice, security, and adherence to Russian legislation. These principles govern the development, deployment, and operation of qualifying AI systems.
The law introduces classifications for sovereign and national models, linked to Russian jurisdiction. A sovereign model must be developed by a Russian legal entity and hosted on data centers located within the country. Its creators must maintain the capability to reproduce the entire development process, including training and original parameters. A national model adheres to similar ownership and localization criteria but may include foreign software components released under open licenses, provided Russian entities retain necessary control and operational capacity.
Legal Designations for Domestic AI Models
The government may provide assistance to developers involved in creating, deploying, or managing qualifying foundation models. Such support could include access to state-held datasets for training purposes. Authorities may also mandate the exclusive use of sovereign or national models within government information systems and other sensitive environments. Additional rules related to defense, security, public order, and property protection may be established through separate legislation or presidential decrees. The framework assigns responsibility to state agencies for ensuring these requirements are met within their legal authority.
Large digital platforms are subject to a distinct obligation concerning AI-generated audio and visual content. Services with over 500,000 daily users must implement tools enabling users to label such content. This requirement applies to websites, applications, and social media platforms. It does not mandate automatic labeling of all items, but developers and users can agree on how notices are presented through service terms. The focus is on providing an option for creators or distributors of qualifying material to disclose AI-generated content.
Standards for Copyright and Content Disclosure
AI providers must inform users about rights ownership of generated material. They must also clarify access conditions and whether the content can be downloaded or transferred. The legislation separately addresses the use of copyrighted works for machine learning, allowing analysis for extraction, comparison, classification, and pattern recognition when lawful access has been obtained. Training on protected works is permitted if no technical restrictions were bypassed to gain access. These rules tie model training procedures to existing copyright and access regulations.
Most of these provisions are set to take effect on September 1, 2026, following presidential approval and official publication. Regulations concerning domestic model classification, developer obligations, content marking, and intellectual property will come into force on March 1, 2027. Existing systems may continue operations until September 1, 2032, provided they process and store data within Russia. Until the legislation is formally signed and published, it remains an approved bill rather than an enacted federal law in Russia’s legislative process.