Spiritual test for AI: only "moral" models will gain access to government procurement in Russia
The Russian artificial intelligence market is on the verge of an unprecedented experiment: developers of foundational models will be required to confirm not only technical compliance with the law, but also their commitment to "traditional spiritual and moral values." This initiative, actively discussed in late July and early August at the ANO "Digital Economy" platform, radically changes the rules of the game for all industry participants.
Formally, the AI law came into force on July 26, but it is framework-based in nature. Now, key requirements must be spelled out in subordinate legislation, and this is where the main intrigue unfolds. The focus is on introducing two special statuses for large models—"sovereign" and "national." Obtaining either of these will become a ticket to state preferences, including advantages in government procurement.
Verification mechanics: from benchmarks to prompt injections
There are no plans yet to create a separate act with a list of values—this list is already fixed in Presidential Decree No. 809. Instead, specialized benchmarks—sets of tests for evaluating models—will be developed jointly with businesses. The procedure will be two-stage: the developer submits an application describing the architecture and filtering mechanisms, after which an accredited laboratory tests the model's responses and attempts to bypass internal restrictions using prompt injections—essentially, hacking the AI with malicious instructions in the input data.
The experts' conclusion is sent to the government commission on digitalization, and only a positive verdict opens access to state support. An important nuance: not all models on the market will be tested, but only those seeking sovereign or national status. The first areas of implementation will be education and government services.
This entire structure raises serious questions for me. Formalizing "spirituality" in the form of a technical test is an almost impossible task. Business architects of AI projects rightly note that categories such as civic-mindedness or historical memory cannot be directly translated into a technical specification. The only possible approach is procedural: a unified reporting format and reproducibility of results. But here lies the main risk—the "evaluative nature" of the criteria, which creates grounds for selective application.
At the same time, the system itself will be decentralized: several competing laboratories that have passed verification by a government body will be able to issue conclusions. The final decision will remain with the Ministry of Digital Development, while security for state systems will be controlled by the FSB and FSTEC. Major developers already have tools for blocking prohibited requests, so the ministry expects to rely on market developments.
Likely candidates for the role of verifiers include the Institute for System Programming of the Russian Academy of Sciences. However, the package of subordinate legislation is still undergoing interdepartmental coordination, and it is premature to discuss the final structure. Lawyers are already proposing to introduce certificate validity periods and a public registry of conclusions to minimize the risks of subjectivity.
My conclusion: the initiative is ambitious, but its success directly depends on whether truly objective and reproducible evaluation procedures can be created. Without this, the "spiritual filter" risks turning from a tool for protecting values into a mechanism of administrative pressure on the market, which would ultimately hit the industry's innovative development.