Criteria for compliance with legislation and traditional spiritual and moral values will be developed for large AI models in Russia — requirements for the models will be established by the Ministry of Digital Development, the FSB, and the FSTEC. An independent testing laboratory will check the models for compliance with these criteria. IT companies will be required to submit their developments to this laboratory if they aim to obtain the status of a sovereign or national model.
Discussion of the verification mechanism took place at the end of July and beginning of August on the platform of the autonomous non-profit organization "Digital Economy" — where by-laws for Federal Law No. 243-FZ "On Supporting the Development of Artificial Intelligence Technologies in the Russian Federation," adopted on July 26, were being prepared.
What the Status of a Sovereign Model Provides
The law introduces two new statuses — sovereign and national large fundamental artificial intelligence models. To obtain either status, a developer will have to confirm that the model complies not only with Russian legislation but also with traditional spiritual and moral values. The procedure for such confirmation will be established by the government.
The development of the corresponding regulations is currently being undertaken by three agencies — the Ministry of Digital Development, the FSB, and the FSTEC. A first draft of the architecture for these rules is already ready. As stated on July 7 during the first reading of the bill in the State Duma by State Secretary and Deputy Minister of Digital Development Ivan Lebedev, the assessment procedure will include criteria to determine whether a model meets spiritual and moral values and the country's technological development goals.
Which Values Will Serve as the Basis
The list of traditional Russian spiritual and moral values is enshrined in Presidential Decree No. 809 dated November 9, 2022 (as amended on March 4, 2026). It includes:
life, dignity, human rights and freedoms
patriotism, citizenship, service to the Fatherland
strong family and creative labor
priority of the spiritual over the material
humanism, mercy, and justice
collectivism, mutual assistance, and mutual respect
historical memory and the unity of the peoples of Russia
There are no plans to prepare a separate act specifically spelling out spiritual and moral values in relation to AI. Instead, special benchmarks will be developed jointly with businesses, and models will be tested against them.
How the Check Will Work
The initiative to create a testing laboratory can be taken by the Government Commission on Digitalization — upon the proposal of the Ministry of Digital Development. The procedure itself looks as follows:
The developer submits an application describing the model's architecture, information about themselves, and content filtering mechanisms.
The laboratory tests the model's responses using benchmarks.
The possibility of circumventing restrictions, including prompt injections, is checked separately.
Following the check, the laboratory sends a conclusion to the Government Commission.
A positive conclusion will be one of the stages for granting the status of a national or sovereign model, and along with it — access to state support measures.
The law does not introduce universal mandatory checks for all AI models on the market — it applies only to those developments that claim sovereign or national status. The law itself comes into force on September 1, 2026, with the exception of certain provisions.
AI's Opinion
From the perspective of machine data analysis, the Russian model of checking for "spiritual and moral values" finds parallels in Asian experience regulating generative technologies. China, as far back as the mid-2020s, established requirements for developers to use only "lawful data" for training models and to disclose this information to the regulator — essentially the same principle of content control, only formulated through the lens of data protection, not ideology. The difference in terminology is indicative: the Beijing approach appeals to the legal purity of the training dataset, the Moscow one — to compliance with the system of values from Decree No.809.
A technical aspect the article does not reveal: how exactly will benchmarks measure the "priority of the spiritual over the material" in a language model's responses — a criterion poorly suited to formalization, unlike checking data for legality of origin. Will such an assessment remain reproducible and transparent for developers, or will it turn into a closed procedure with an unpredictable outcome?
