Russia Passes AI Regulation Law — Sovereign AI Models Must Be Homegrown

Russia’s State Duma passed a sweeping AI regulation bill on Tuesday, creating a legal framework that splits large AI foundation models into two categories — sovereign and national — both of which must be developed by Russian legal entities.

The law, which cleared its second and third readings in a single session, is designed to establish the legal conditions for developing and deploying large AI models inside Russia while preserving what the government calls the country’s technological independence. The text was published on the Duma’s website the same day.

At the heart of the legislation is a two-tier classification system. A “sovereign AI foundation model” must be fully controlled by a Russian legal entity at every stage of its lifecycle, from development through deployment. All data centers running the model must be physically located inside Russia. A “national AI foundation model” is also required to be built by a Russian legal entity, but it can incorporate foreign components released under open-source licenses. Both categories must undergo compliance checks to confirm they align with Russian law and what the legislation describes as the country’s traditional “spiritual and moral values.”

The government has the authority to mandate the use of sovereign or national AI models in state information systems and other sensitive areas. The president will approve the national AI development strategy, while the government coordinates development and implementation efforts and decides on support measures.

On content labeling, the law takes a lighter touch than some observers expected. It does not require all AI-generated content to be explicitly marked. Instead, websites, apps, and social media platforms with more than 500,000 daily active users must offer their users the option to label AI-generated content. The labeling format will be worked out between AI service developers and users.

On copyright, AI service providers must inform users about the copyright status of AI-generated content, along with access terms and download methods. The law also permits the use of copyrighted materials for AI model training — as long as the material was obtained legally and without bypassing technical protections — and does not require consent from the rights holder.

The timeline is staggered. Most provisions take effect on September 1, 2026. But the rules around sovereign and national model requirements, developer obligations, content labeling standards, and intellectual property rules for model training won’t kick in until March 1, 2027. Existing AI systems that already store data inside Russia get a transition period running through September 1, 2032.

The bill still needs approval from the Federation Council, Russia’s upper house, and a signature from the president before it becomes law. According to sources familiar with the matter, Sberbank’s AI model already meets the sovereign standard, while Yandex’s model qualifies under the national classification.