Roskomnadzor uses AI to block sites

This year, Roskomnadzor is going to start maintaining a register of prohibited information using artificial intelligence (AI). The technology will work on the basis of a system through which texts on the Internet are already analyzed and classified. The integration of AI technologies is designed to reduce costs and establish “non-obvious connections.” In two years, the department expects to use such technologies to maintain a register of personal data operators. But this task, experts believe, will be more difficult to automate.

Kommersant got acquainted with the new version of the passport of the Roskomnadzor digital transformation program. According to the document, the department plans to create and maintain a register of blocked sites using AI starting in 2024. This is stated in the description of the work related to the unified information system of Roskomnadzor (EIS, which also combines registers of licenses, media, permits) and the information system for monitoring Internet resources (IS MIR). In 2023, based on passport data, the register of prohibited sites was maintained without the use of AI.

IS MIR, as follows from the tender documentation of 2021, is designed to track texts with prohibited information, classify them by nature (neutral, negative or positive opinion of the author) and search for reprints. Last year, the Federal State Unitary Enterprise “Main Radio Frequency Center”, subordinate to Roskomnadzor, announced a tender for the integration of the MIR IS with other systems, including the Oculus IS, designed to search for images and symbols (see “Kommersant” dated September 30, 2023).

In 2023, the document says, Roskomnadzor identified illegal content on the Internet three hours after publication. In 2024, the figure is planned to be reduced to two hours, and by the end of 2026 – to one. The service also plans to improve its efficiency: in particular, to reduce the rate of erroneously identified signs of violations in the media from 20% in 2023 to 10% in 2026. Roskomnadzor did not respond to Kommersant’s request.

The use of AI, rather than specified text processing algorithms, will reduce the amount of human resources to maintain the system, says Innostage product manager Evgeniy Surkov. To do this, you need to train the AI ​​model on a sample of materials from a verified and operator-controlled system. The model can then operate with minimal human intervention, he said.

AI will make it possible to “identify complex contextual connections between text fragments, find hidden patterns and associations,” notes Just AI product manager Alexey Borshchov. The technology can also be used to structure databases: “But it is difficult to name typical examples: usually the knowledge base is not formed using AI, but rather is integrated with it.”

The head of the investigation department at T.Hunter, Igor Bederov, believes that “even after two years, the share of detected prohibited content that will require additional human moderation is unlikely to be below 60%.” According to him, it is difficult to assess the speed of the system from the outside – analysts base it on the timing of the content blocking itself: “In 2022, blocking individual sites with prohibited information took up to 50 days.”

In 2026, according to the passport, Roskomnadzor also intends to fill out the register of personal data operators with the help of AI. Aleksey Boyko, an analyst at the specialized Telegram channel abloud62, admits that it is possible to automatically identify personal data operators using open information: “But there is a risk that at first such AI will produce false detections.”

Identifying operators is fraught with difficulties, emphasizes Natalia Tylevich, CEO of data analysis systems developer Social Laboratory: “Even if an operator has published legal documents in a form convenient for automatic extraction, their “reading” requires deep semantic analysis. But, for example, an online store can only display them during the user registration or ordering process.” In the latter case, she clarifies, solution creators will have to find a way to automatically register on sites whose administrators usually block such actions.

Yuri Litvinenko