Russia Adopts Standards Set to Give AI Development a New Boost
Data sharing between organizations is set to become easier, while Russian technologies and platforms could attract interest in overseas markets.
How to Assess Dataset Quality
Russia has approved a series of three preliminary national standards, or PNSTs, governing the creation and use of synthetic data – artificially generated datasets that reproduce the statistical properties of real data without being directly linked to specific individuals. The standards are PNST 1064-2026, “Data Synthesis. General Provisions”; PNST 1065-2026, “Data Synthesis. Data Synthesis Process Architecture. Synthesis Methods”; and PNST 1066-2026, “Data Synthesis. Description of Data Synthesis Process Results. Quality Assessment Methodology.” Rosstandart (Federal Agency for Technical Regulating and Metrology) has already approved the standards, which are scheduled to take effect Sept. 1, 2026.
The nonprofit Assotsiatsiya bolshikh dannykh (Big Data Association) developed the standards with participation from the nonprofit National Technology Center for Digital Cryptography. Work began in 2025.
The documents establish common rules for generating synthetic data, define the architecture of the systems used to produce it and set methods for assessing quality and privacy. Synthetic data can help organizations build large datasets when real-world information is unavailable, contains personal or commercially sensitive information, or is too expensive to collect. For the first time in Russia, companies will have a common methodological framework for assessing the quality and safety of synthetic datasets.
Data Sharing Will Get Easier
Russian developers will now have access to a broader data foundation for AI. Such datasets could be particularly useful in fields that need large volumes of machine-learning data while also facing strict confidentiality requirements, including banking, telecommunications, government, health care and industry. Standardization could also make it easier for organizations to exchange datasets by establishing uniform requirements for data generation, process documentation, and quality and privacy assessments.
Technologies for generating and securely sharing data, along with enterprise platforms that prepare datasets for AI, could eventually become export products as well. At least, the draft standards stated that no direct international or regional equivalents to the Russian documents existed when they were being prepared. That could give Russia an opportunity to promote its own approach to synthetic-data standardization, including through technology partnerships with BRICS countries.
From Draft to Testing in 18 Months
Back in 2023, Sber developed SyntData, its own ecosystem for generating synthetic data to train AI and work with confidential datasets. In 2024, the service was added to Russia’s Unified Register of Russian Software. By January 2025, Assotsiatsiya bolshikh dannykh had announced work on a preliminary data-synthesis standard together with Sber and other industry participants. That September, Assotsiatsiya bolshikh dannykh and the Academy of Cryptography held a workshop covering synthesis-system architecture, differential privacy, GAN and VAE generators, and methods for assessing synthetic-data quality.
By June 2026, synthetic data was already being discussed as part of the broader data-economy agenda. In July, Rosstandart approved the PNST 1064–1066 series. Work that began with a draft standard in 2025 has thus moved into formal testing of the regulatory documents.
Real Data Is Not Going Away
The standards are to be tested in practice over the next three years. PNST 1064 and PNST 1066 are valid through Sept. 1, 2029, after which experience accumulated across industries could inform revisions to the documents or the conversion of their requirements into permanent national standards.
The greatest impact is likely in fields where a shortage of accessible, high-quality data limits AI adoption. Synthetic data, however, should not be viewed as a complete replacement for real-world data: model quality depends directly on how accurately an artificial dataset reproduces real patterns without carrying over errors or biases from the source data. That is precisely why one of the three standards is dedicated to methods for assessing quality and privacy.
