Russia adopts law on "sovereign" artificial intelligence

Russia now has its own version of artificial intelligence regulation. The State Duma passed the law in its second and third readings on July 8, a day after the first reading, and the Federation Council approved it in the following days. The bill is now on Vladimir Putin's desk and will take effect on September 1, 2026, if the president signs it.

Two model categories

At the heart of the law is the division of large foundation AI models into two categories. A "sovereign" model must be developed and fully controlled by a Russian legal entity throughout its entire life cycle, with its infrastructure running exclusively on servers located in Russia. A "national" model must also be created by a Russian company, but it may use foreign open-source components, as long as the data and servers remain within the country.

According to TASS, among domestic players Sberbank meets the criteria for a sovereign model, while Yandex qualifies as a national model. The Russian government has been granted the authority to designate cases in which only sovereign or national models may be used in public administration and sectors deemed sensitive, effectively shutting out systems such as ChatGPT, Gemini and Claude from those areas.

Traditional values as a legal requirement

The law establishes the principle of "taking into account and respecting traditional Russian spiritual and moral values" as one of the core regulatory criteria for AI models operating in the country. The provision survived into the final version of the legislation despite earlier objections from parts of the tech industry, which worried that a requirement phrased this way would be difficult to translate into concrete, measurable technical criteria.

Earlier drafts of the bill, reported on by the Russian outlet Meduza back in June, sparked disputes over whether sovereign models would be allowed to train on data from outside Russia. The final version softened that restriction, permitting training on non-domestic data while keeping the requirements for Russian entity control and in-country infrastructure location.

Copyright and content labeling

Beyond dividing models into categories, the law also addresses two other issues that are contentious worldwide: copyright in model training and the labeling of AI-generated content. The rules allow copyrighted material to be used for training models without rights holders' consent, putting Russia in direct opposition to the direction courts in the US and Europe have been heading after high-profile settlements, such as Anthropic's recently approved $1.5 billion settlement with book authors.

Operators of platforms with more than 500,000 daily users will be required to provide the technical means to label AI-generated content, though labeling itself remains voluntary, with its format to be set out in agreements between AI providers and users. That is a considerably lighter-touch approach than in the European Union, where mandatory AI content labeling takes effect on August 2.

The law is intended to establish a legal framework for the rapid growth and deployment of large foundation AI models in Russia - from the explanatory note adopted by the Federation Council

Russia's version of technological sovereignty

The term "sovereign AI" has circulated in the global debate for months, driven largely by calls from chipmakers and experts for open ecosystems independent of a handful of US suppliers. Russia's law gives the concept an entirely different, legally binding meaning: rather than a call for openness, it functions as a mechanism for state control over which models may be deployed in government and strategic sectors, and as a tool for forcing content to align with the official ideological line.

For Polish readers, the contrast with the EU AI Act, whose successive provisions are taking effect in Poland this year, is stark. While Brussels and Warsaw focus on risk assessment of systems and transparency toward users, Moscow is building a legal framework centered on controlling the origin of technology and compliance with state values, while simultaneously loosening copyright protections for creators.

The law does not yet contain a detailed catalog of penalties for violations, leaving Russian regulators broad room to work out implementing regulations before the main provisions take effect in September. Key requirements concerning sovereign and national status, developer obligations and the use of intellectual property will only apply from March 2027, giving Russian AI companies more than six months to adapt to the new rules.