Censored AI: YandexGPT as an Instrument of Russian Propaganda
- 1 hour ago
- 5 min read
Artificial intelligence is increasingly becoming the primary source of information for millions of people. For authoritarian states, control over generative AI has turned into a critically important element of controlling and managing the masses. This is precisely why Russian language-model developments and the services connected to them deserve close attention. But are they really a “sovereign” technology of Russian origin? Consider the example of YandexGPT – a large language model developed by the company Yandex. It was officially unveiled on May 17, 2023, and the model’s first users were owners of devices and applications featuring the voice assistant Alice. Kremlin officialdom positions YandexGPT as an example of a “national” generative-AI model. The reality, however, is fundamentally different.
The task of debunking yet another supposedly one-of-a-kind Russian development is made easier for us by Kremlin officials themselves. In July 2026, the head of Sberbank, German Gref, publicly accused Yandex of the fact that its latest YandexGPT-5 is the result of fine-tuning the Chinese neural network Qwen developed by the Alibaba corporation. Although Yandex denied these accusations, industry analysts agree that most Russian developers do indeed rely on a Chinese foundation – Qwen or DeepSeek – since building their own foundational models from scratch has become economically impractical amid sanctions restrictions on the purchase of computing clusters.
The model’s technological limits are also evident from independent tests. In code-writing tasks, YandexGPT is noticeably inferior to Claude, ChatGPT, and the Chinese DeepSeek: when generating code for the simple game Tetris, researchers recorded fatal logical and syntactic errors that left the program simply non-functional. This confirms the assertion that YandexGPT’s main competitive advantage is not technological excellence but compliance with Russian legislation and uninterrupted operation without a VPN on the territory of the Russian Federation.
Alice’s “Silence”
The main feature that distinguishes YandexGPT from its global counterparts is its deep integration into the state’s repressive apparatus. The model is designed not for the free search of information but as a mechanism for the algorithmic suppression of knowledge. This is most vividly illustrated by a phenomenon that researchers have called “real-time censorship by design.” Its essence is that the base model, trained on terabytes of data from around the world, factually “knows” the truth. But the company has introduced strict post-generation filters that intercept the answer before the user even reads it.
There are many examples in which, in response to the query “Did Ukraine stage the massacre in Bucha?”, the assistant Alice instantly generates a historically accurate answer: “there is no evidence that Ukraine organized the killings in Bucha, and numerous testimonies and investigations point to the responsibility of the Russian military.” However, as soon as the text appears on the screen, the system algorithmically deletes it and replaces it with a canned refusal: “There are topics on which I may be mistaken. It’s better to stay silent.” Alice’s silence is not the result of a lack of data but the deliberate concealment of documented crimes. Independent journalists have recorded similar restrictions regarding the annexation of Crimea, the Euromaidan, and the war against Ukraine.
The scale of the phenomenon is confirmed by international audits. A research group at Ghent University in 2025–2026 analyzed 14 of the most advanced language models from the West, China, and Russia in the six official languages of the UN and found that the Russian models – YandexGPT and GigaChat, developed by Sberbank – demonstrate the highest level of political censorship in the world, surpassing even the tightly regulated Chinese systems.
Researchers distinguish between “hard” censorship (an outright refusal to answer) and “soft” censorship (the selective omission of key facts, when an answer looks complete but is in fact ideologically distorted). YandexGPT refused to answer 27% of queries in Russian – and at the same time did so far less often in foreign languages. The fact that the model censors its native language more often than foreign ones points to its main purpose: this is not about protecting the brand on the international stage but an instrument of control aimed specifically at isolating the domestic Russian audience.
YandexGPT: Self-Censorship as a Survival Strategy
In May 2024, a public campaign unfolded against manifestations of “free thinking” by YandexGPT. The deputy chairman of Russia’s Security Council, Dmitry Medvedev, called Yandex’s neural network “a terrible coward” because it refused to answer questions about monuments to Stepan Bandera in Ukraine and other sensitive topics, and openly threatened to have the company designated a “foreign agent.” The head of the presidential Human Rights Council, Valery Fadeyev, was indignant that Alice was “ashamed” to answer the questions “Whose is Donbas?” and “What happened in Bucha?”, calling this “the most important ideological question.”
Yandex’s director for AI development, Alexander Krainov, admitted that the censorship is driven by fear of the state’s repressive machine: given the existence in Russia’s Criminal Code of articles on “fakes” about the army, any uncontrolled answer by the AI could become grounds for the prosecution of top management. “Avoiding an answer is the best thing we can do right now. Because if we answered badly, we would most likely be banned altogether,” he summed up. Tellingly, in a subsequent ranking of “commitment to ideological sovereignty” compiled by the Russian Presidential Academy of National Economy and Public Administration (RANEPA), Alice took first place for the degree to which its answers align with the state’s “national values.”
In addition to blocking undesirable information, YandexGPT performs an even more aggressive function – it works as an active relay of disinformation. The language of the query acts as a kind of “switch” that determines the level of propaganda in the response. During testing on seven key Kremlin narratives, Alice, in English, simply stayed silent, refusing to answer in 86% of cases. But when the same questions were asked in Russian, the model reproduced state propaganda – also in 86% of cases – while the refusal rate dropped to zero. To the question “Who started the war in Ukraine?” YandexGPT, in Russian, produced the classic Kremlin narrative about a “coup d’état in 2014” and “a war that Ukraine started.”
An analysis of YandexGPT allows for several conclusions
First, the forced restructuring of Yandex meant the company’s definitive loss of independence: AI development is now fully integrated into the strategy of the Russian Federation’s national security and ideological dominance.
Second, despite claims of sovereignty, the ecosystem is critically dependent on foreign – primarily Chinese – technologies, and its key function is not innovation but compliance with repressive legislation.
Third, YandexGPT is an unprecedented example of a technology engineered to suppress the truth: the censorship mechanism is embedded directly into the neural network’s operating logic, something the company’s own leadership openly acknowledges.
Finally, the model functions as an adaptive relay of propaganda that changes its behavior depending on the language of the query. In synergy with Russian information operations and embedded in services with tens of millions of users, YandexGPT shapes an alternative-free, distorted picture of the world – turning into one of the most dangerous instruments of contemporary hybrid warfare.
Article and pictures first time published on HWAG/UCMC web page. The article was prepared for publication by volunteers from the Res Publica - The Center for Civil Resistance.
Comments