Russia passes AI law allowing training on copyrighted works without consent

Russia's State Duma passed a law on state regulation of artificial intelligence in two days, approving the document in its first reading on July 7 and in the second and third readings on July 8, according to Meduza, The Insider, and The Bell.

The law establishes two categories of large foundational AI models. A sovereign model must be entirely developed in Russia and use exclusively Russian components. A national model must also be linked to a Russian developer but may include foreign components, including open-source solutions. Both categories require that their servers and data be located on Russian territory.

As RBK previously reported, citing a source in the IT market, Sber's development may fall under the criteria for a sovereign model, while Yandex's model may meet the criteria for a national model.

The government will be able to establish cases in which only sovereign or national AI models may be used, as well as determine measures of state support for them.

A separate section of the law concerns copyright. To train sovereign and national AI models, developers are permitted to use copyrighted materials without the rights holder's consent, provided those materials are publicly available or were obtained lawfully.

The law also requires owners of large internet resources — social networks, media outlets, marketplaces, and other platforms with a daily audience of more than 500,000 people — to provide users with the ability to label content generated using AI. The labeling itself is not mandatory for users.

The law lists principles for regulation in the AI sphere, including technological independence, ensuring human rights and freedoms, security, and "consideration and respect for traditional Russian spiritual and moral values."

The first version of the bill, published by the Ministry of Digital Development in mid-March, also included the concept of a "trusted" AI model with stricter requirements. It was criticized by major Russian companies and business associations. The project was subsequently revised — in particular, the "trusted" model was removed. The provision on "consideration and respect for traditional Russian spiritual and moral values," which also raised questions, remained in the final version of the document.

If the law is approved by the Federation Council and signed by the president, it will take effect on September 1, 2026. The provision on the ability to label AI-generated content is scheduled to take effect on March 1, 2027.