Government softens AI regulation bill after business criticism; requirements for data sources and developers were removed from it

The Russian government has softened the rules of the bill on the regulation of artificial intelligence after criticism from business. RBC reports this with reference to a representative of the Deputy Prime Minister’s office, Dmitry Grigorenko.

The draft law “On the fundamentals of state regulation of the application of artificial intelligence technologies in Russia” was published in March. The initial version caused criticism from businesses: companies warned that new requirements could increase the cost of implementing AI by 20–40%, and slow down the launch of products to the market by one and a half to two times.

The document envisaged the introduction of three categories of AI models: “sovereign”, “national” and “trusted”. It was planned to allow only “trusted” models for access to government information systems and critical information infrastructure objects.

After finalization, the requirement for data sets for training models was removed from the bill. In the first version, it was assumed that “sovereign” and “national” models should be trained only on data created in Russian by Russian individuals and legal entities. Business criticized this: market participants felt that the necessary Russian-language data in open sources was not enough, and the restriction could reduce the quality of the models.

The document also excluded the requirement that only Russian citizens be involved in the development and training of “sovereign” and “national” models. Now, to do this, it is enough to open a Russian legal entity and confirm the compliance of the AI ​​model with Russian legislation and “traditional spiritual and moral values.”

In addition, the new version removed the requirement for AI services with an audience of more than 500 thousand users to register as an organizer of information dissemination. This status implies the obligation to establish systems of operational investigative measures (SORM), through which intelligence services can gain access to correspondence and user data.