Russian AI to be checked for compliance with traditional values — tests to be conducted by independent laboratories

According to Vedomosti, Russian AI models may start being checked for compliance with legislation on traditional spiritual and moral values. The assessment will apply to neural networks claiming the status of national or sovereign. The norms obliging to check AI models for compliance with traditional values have not been adopted yet; they are being discussed in the Ministry of Digital Development. According to the AI law adopted on July 26, two special statuses appeared in Russia for large models — national and sovereign. They allow the developer to apply for state support, access to data of state information systems, and priority use in certain areas. To obtain the status of a national or sovereign AI model, companies need to confirm that the model complies with Russian legislation and traditional spiritual and moral values. Authorities and businesses propose creating special benchmarks to evaluate AI responses. AI model developers will test neural networks independently using a unified methodology, after which independent laboratories will study the results and, if necessary, conduct additional checks. Specialists will check whether it is possible to force the AI to bypass restrictions and produce a response that does not correspond to traditional values. Independent laboratories that have passed accreditation will be allowed to check AI models; requirements for organizations are being developed together with business. To pass the examination, the developer must provide a description of the model architecture, information about its creator, and data on content filtering mechanisms. Based on the results, the laboratory will prepare a conclusion on compliance. The final decision on assigning the status of national or sovereign model remains with the Ministry of Digital Development. Experts warned that AI assessment may be subjective, as it is unclear what specific requirements neural networks must meet. Lawyers also noted the risk of subjective evaluation and suggest assigning checks to several independent laboratories and allowing their decisions to be challenged.