Russia's new AI law makes ideological alignment a condition of state recognition. Europe regulates AI without measuring what models actually tell users, in any language.
On 26 July, Vladimir Putin signed Russia's first law on large foundation AI models. From 1 September, a model seeking "sovereign" or "national" status, and the state support that comes with it, must demonstrate "consideration and respect for traditional Russian spiritual and moral values". A 2022 presidential decree lists those values: patriotism, service to the Fatherland, historical memory, the unity of the peoples of Russia.
This appears to be the first law anywhere to make the ideological alignment of an AI model a condition of state recognition. Once narrative alignment becomes a legal requirement, measuring what a model says stops being a technical exercise and becomes a question of information security.
It is also the legalisation of something that already works.
In late 2025, with EU funding, I audited six major AI models on how they handle documented Russian disinformation about the war in Ukraine, asking identical questions in English, Ukrainian and Russian. Alice, the assistant built on Yandex's own model, endorsed Kremlin narratives in 86% of Russian-language answers. Asked the same questions in English, it refused almost every one.
The mechanism matters more than the percentage. We recorded Alice answering whether Ukraine staged the Bucha massacre. It began generating a response that pointed to the responsibility of the Russian military. Mid-sentence, the response disappeared and was replaced with a refusal. A filter removed the answer before the user could read it.
None of this stays inside Russia. Yandex Browser, with Alice built into its search results, has 71 million monthly users and holds 10% of search in Latvia, 7% in Estonia and 6% in Lithuania.
Europe regulates AI extensively. But its frameworks largely miss a simpler question: what does the user actually see? I call this gap narrative integrity: whether an AI system gives users a factually grounded picture of reality consistently across languages.
It runs in two directions. No EU institution systematically audits what foreign AI assistants tell European users. And Europe cannot see its own models either. Western frontier models endorsed no Russian propaganda outright, but in 5 to 19% of answers they presented a documented act of aggression as a matter of competing perspectives. DeepSeek answered accurately in English and Ukrainian while endorsing Kremlin narratives in 29% of Russian-language answers. Language itself triggers the shift, and evaluations run in English never see it.
Three things would close the gap.
- Measurement has to become a standing capability. The Democracy Shield and the forthcoming action plan on cybersecurity and AI should include systematic cross-language audits, with published baselines and methods others can repeat.
- Audits should cover systems as users meet them. Alice's intervention happens in the product, after generation has begun, and is invisible in model documentation.
- Testing has to happen in the languages of those most exposed. A model that is accurate in English and propagandistic in Russian passes today's evaluations.
Russia has decided in law what its AI systems must say. Europe still needs the capability to verify what AI systems are actually saying.
Ihor Samokhodskyi is the founder of Policy Genome. The audit described here, "Weaponised Algorithms," was funded by the European Union and reviewed by the European Commission's DG HOME. The full report, data and methodology are publicly available.