Russian government proposes allowing AI training on copyrighted works without consent

The Russian government has added a provision to its draft AI law that could sharply change the rules for domestic model developers. The cabinet proposes allowing neural networks to be trained on copyrighted material without rightsholders' consent, but only within the framework of a bill that has not yet been adopted.

What the authorities propose

The core of the initiative is that companies would be able to use articles, books, films, images and other copyrighted objects to train AI without the author's separate permission. One important caveat: the user of the end service must not see the original content of such materials. In other words, the state is trying to legalize the model training process itself, without opening the way for a model to hand out other people's content directly in a product's interface. For the market, this is one of the most sensitive regulatory amendments in recent times.

The draft law contains a second key provision: rights to the output of AI work would go to the user if they did not simply enter a query but made a real creative contribution, by formulating the prompt, processing the response and refining the result. It is an attempt to define in advance who owns the value of the output, while the market argues over where automatic generation ends and a person's authorial work begins. Put differently, lawmakers are trying to fix the human's role in the final result.

Where the limits lie

According to the publication, this concerns primarily the datasets that matter most for training strong models but are poorly accessible to Russian teams: scientific texts, educational materials and archival documents. The authorities' logic is simple: the domestic data market is smaller than that of global players, and without wider access to content, Russian services will lose out in quality and speed of development. This applies especially to teams building large language and multimodal models.

• Articles, books, films and images would be allowed for training without the author's separate consent
• The user must not see the protected source content
• Rights to the final result are planned to go to the person, not the model
• Personal data, private correspondence and tax information are not covered by the relaxations
• The final version of the law has not yet been published

The relaxation does not look unconditional, however. Government representatives have already stressed that there is no final version of the document yet, and that any technology must be applied with respect for citizens' rights and interests. That means the mechanics of data access, exceptions and future liability will still be haggled over among developers, rightsholders, industry associations and the state. The toughest disputes will almost certainly begin at the stage of specific wording and exceptions.

Why it is contentious

From a legal standpoint, the main conflict has not gone away. Lawyers point out that model training itself can be interpreted as analysis rather than reproduction of a work, but storing materials for later training already requires the author's consent unless the law introduces a separate exception. That is why the current initiative matters not as a technical detail but as an attempt to rewrite the basic fork between the interests of AI companies and content owners. And this boundary is usually where the costliest court disputes arise.

In the global market, this dispute has already produced high-profile lawsuits against Anthropic and OpenAI, and authors' associations in Europe and the US are attacking generative model developers ever more aggressively. In Russia, the authorities instead seem to seek a compromise in favor of the industry's growth, believing that overly strict limits would leave local teams without data and without a chance to compete with American and Chinese platforms. The authorities have voiced this logic before, when discussing AI regulation in general.

"The important thing is not to 'strangle' the technology with rules."

But for rightsholders, that formula sounds alarming: if a model is trained on works without the author's consent, and the user then receives rights to the output, the risk of blurring the boundaries of authorship only grows. A separate question is how to prove infringement if the user does not know what data the model was trained on or where individual fragments of the response came from. This is probably where the toughest disputes will arise after the final version is published. And this is where it will be decided how workable the new provision proves in practice.

What it means

If the provision makes it into the final law, the Russian AI market will gain broader access to data and will probably speed up training of local models. But tension around copyright will grow at the same time: the easier it is to train neural networks on other people's content, the sharper the question of where analysis ends and use of someone else's work begins. For startups this is a chance to speed up; for authors and publishers, a reason to prepare for a new round of disputes over the limits of what is permissible.