AI bill increases liability for owners of online services and neural networks

New provisions have been added to the Russian draft law on artificial intelligence that expand the obligations of online service owners and AI developers. The document more precisely describes the rules for using and training AI, establishes marking of synthesized content, and strengthens liability for its unlawful application.

What Changes in the Draft

The amendments make regulation noticeably more practical. While previous discussions around the draft law mainly focused on basic definitions and general principles, the emphasis is now shifting toward practical AI use in digital services. This concerns not only the models themselves, but also platforms that embed them in search, support, text generation, image generation, and other user scenarios. For the market, this signals that service owners will be expected to implement not just formal rules, but actual control measures.

Separate clarifications address issues of model training and handling of synthesized content. The requirement to mark materials created or significantly altered using AI is effectively being established as a mandatory part of the future regulatory regime. This is important for media, marketplaces, educational platforms, and any products where users might not distinguish between machine-generated and human-created work. As of April 2026, many companies were using such tools freely, but the new norms clearly guide the market toward stricter accounting.

• Rules for using and training AI systems are clarified
• Responsibility of neural network owners for unlawful application is strengthened
• Mandatory marking of synthesized content is established
• Online services that embed AI in their products may also face stricter requirements

In practice, this means a shift from the idea that "the model is neutral, only users are responsible" to a model of shared responsibility. If a service provides access to generation, publishes results, or uses AI in public scenarios, it may be expected to have clear rules, warnings, and mechanisms for rapid response. Even if specific procedures are still being refined, the direction is already clear: the state wants to see a controlled environment, not a fully self-regulated market.

Where the Risks Are for the Market

The main question for business is not the fact of regulation itself, but the cost of compliance with new rules. Large platforms can usually afford moderation, legal expertise, content verification processes, and separate security teams. Small businesses often lack such resources. If a service uses ready-made external models for generating descriptions, images, customer responses, or advertising materials, it will probably have to additionally think about marking, storing generation traces, and rules for responding to misuse. This is why market participants warn of possible pressure on small companies that by 2026 have already grown accustomed to quickly incorporating AI tools without complex compliance infrastructure.

The broader the interpretation of service owner responsibility, the higher the risk that startups and small teams will slow down the launch of new features or abandon certain scenarios altogether. This is especially sensitive for products where users themselves create content and the platform merely provides them with generative capabilities.

That said, the draft law should be read more as a direction of movement than as finally fixed rules. For the market, it is now critical how precisely the final version will distribute responsibility among the model developer, the service owner, and the direct user. This determines whether new requirements will become an understandable working standard or a barrier for those who cannot maintain expensive control infrastructure.

What This Means

The AI market in Russia is moving from free experimentation to formal obligations and verifiable procedures. For major players, this will likely become another compliance zone, and for small businesses, a reason to review in advance how generative tools are used, where control is needed, and where machine-generated content will need to be explicitly marked. For services with user-generated content, this already looks like a signal to prepare processes in advance rather than wait for the final text of the law.