Digital exam on spirituality: how Russia will certify AI for government procurement
The Russian artificial intelligence market is on the brink of tectonic changes: access to government procurement and support measures will now depend on successfully passing a "spiritual and moral examination." This initiative, actively discussed in late July at the ANO "Digital Economy" platform, involves creating a system of independent testing laboratories that will assess whether models comply not only with the letter of the law but also with traditional values.
This involves introducing two new statuses for large foundational models — "sovereign" and "national." To obtain one of these, a developer must submit an application, an architecture description, and request filtering mechanisms to an accredited laboratory. Then the most interesting part begins: experts test the model's resilience to prompt injections — attempts to "hack" it through malicious instructions in input data.
Architecture of the new certification system
The key intrigue remains the content of the tests themselves. There are no plans yet to draft a separate act with a list of values — instead, special benchmarks will be developed jointly with businesses. The procedure is to be finalized in a government decree, with the final conclusion issued by the digitalization commission. A positive verdict grants access to preferences in government procurement, making this procedure critically important for all major market players.
The system will be two-tiered: first, the developer independently tests the model using a risk-oriented methodology, then an accredited laboratory evaluates these findings and may selectively verify resilience. The final say will rest with the Ministry of Digital Development, while security for government systems will be overseen by the FSB and FSTEC. Notably, not all models will be checked — only those seeking status positions, and the first implementation areas will be education and government services.
The question of who exactly will become the "spiritual censor" remains open. Businesses insist on a plurality of laboratories and an appeals board to avoid "evaluative" criteria. The Institute for System Programming of the Russian Academy of Sciences is being touted as a potential examiner, while experts warn that abstract categories such as civic consciousness and historical memory are extremely difficult to formalize in a technical specification. The package of subordinate regulations is still undergoing interagency coordination, so the final configuration of the system will be determined in the coming months.
My view: This is a landmark precedent — the regulator is attempting to embed value orientations into technical AI certification. For the market, this is a dual challenge: on one hand, a clear entry barrier emerges; on the other, there is a high risk of subjectivity. The ability to pass this examination will become a new competitive advantage, and I expect that the battle over benchmarks will be no less fierce than the battle over the models themselves.