Russian AI Bill Softened After Criticism: What Changed and Why It Matters

Between April and May 2026, the Russian government finalized its bill on the fundamentals of artificial intelligence regulation, softening several key requirements following widespread criticism from the business community. The changes affect model training rules, sovereignty criteria, and the responsibilities of online platform owners. We break down exactly what has changed and how it will impact AI developers and users.
What Was Removed from the Bill
The initial version of the document required that only Russian-origin data be used for AI training. Market participants pointed to a shortage of high-quality Russian-language datasets and the risk of reduced model accuracy. In the new version, this restriction has been lifted—developers are now permitted to work with any available data.
The requirement that sovereign and national models be developed and trained exclusively by Russian citizens has also been repealed. Now, to obtain such status, it is sufficient for the development to be conducted by a Russian legal entity and for the model to comply with legislation and established value-based criteria.
The provision classifying services with an audience of over 500,000 users as "organizers of information dissemination" was excluded from the document. This exempts them from the obligation to install user data access systems and simplifies the operation of online platforms.
Who Is Now Responsible for AI Content Labeling
Website and online platform owners have been exempted from the duty to independently verify whether user-uploaded content created with the help of artificial intelligence is labeled. This task is transferred to the owners of AI services, and the government will define specific cases where labeling is mandatory.
In the initial version, online resources were required to monitor the presence of labels on photos, videos, and texts created by neural networks and to remove unlabeled materials. This requirement has been excluded from the new version. The government will determine specific categories of AI content for which labeling will become mandatory: news and political materials, advertising, and educational products.
For owners of websites, marketplaces, and social networks, this means a significant reduction in regulatory burden. They will not have to create separate systems to verify the origin of every image or text or make blocking decisions based on probabilistic algorithm estimates.
Access to State Data and the Trusted Model Registry
The new version provides for the possibility of training national and sovereign neural networks on state data, but only after coordination with the FSTEC and the FSB of Russia. The document also establishes liability for AI operators and owners for ensuring the information security of their models.
The bill now defines an AI as "trusted" only if the model is included in the corresponding registry. The procedure for maintaining the registry of trusted AI and including models in it will be established by the government. Previously, the requirement for registry inclusion was intended only for models used in critical information infrastructure facilities and state information systems.
The government will also be able to establish cases for the mandatory use of sovereign and national AI in the banking sector in coordination with the Central Bank. Criteria for inclusion in the trusted model registry are planned to be defined by separate acts of authorized bodies.
What This Means for Developers and Users
The softening of requirements lowers barriers for developers: they can use international datasets and engage foreign specialists, provided they register a Russian legal entity. This is especially important for startups and companies working in healthcare, where AI use is already integrated into clinical practice.
Market participants note that the initial version could have led to a 20-40 percent increase in AI implementation costs and slowed product time-to-market by a factor of one and a half to two. The new version removes some of these risks while maintaining requirements for security and content labeling.
For users, the main change is the right to pre-trial appeal of decisions made by authorities using AI, and the possibility of receiving compensation under general civil law rules in the event of harm caused by its use.
If passed, the bill will come into force on September 1, 2027. The Ministry of Digital Development reports that the document is being finalized in cooperation with the industry and relevant agencies, and it is too early to discuss the final version and specific details.
If you need to test the capabilities of modern models—from GPT and Claude to Gemini—try GEMERA AI. Access is available via a Telegram bot and web dashboard, with no VPN or complex settings required.