In May 2024, Dmitry Medvedev – former Russian president, current deputy head of the Security Council, and reliable barometer of Kremlin rage – picked a fight with a smart speaker.
The target of his fury was Alice, the AI voice assistant produced by Yandex, Russia’s tech giant. Alice had refused to answer a question about Ukrainian monuments to Stepan Bandera, the wartime Ukrainian nationalist leader. By declining to engage with the topic, Alice was, in Medvedev’s view, undermining Russia’s central narrative that Ukraine is a country run by Nazis. He called Yandex’s managers “terrible cowards” and hinted they should receive the designation of foreign agent – Russia’s label for dissidents and enemies of the state.
Alice’s crime was not incompetence. It was caution. She had been programmed to avoid politically sensitive topics. And in Russia in 2024, the list of politically sensitive topics is so long, and the penalties for getting the answer wrong so severe, that the safest strategy for a Russian AI is often to say nothing at all.
This is the central paradox of Russia’s artificial intelligence ambition: the Kremlin wants to build world-class AI, but has created an environment in which world-class AI is almost impossible to build.
Putin’s Ambition
Vladimir Putin has been publicly committed to AI dominance for years. “Russia must become a world leader, not only in the creation, but also in the penetration of artificial intelligence into all spheres of our lives without exception,” he declared at a government-linked AI conference in Moscow in December 2024.
Five years ago, this might have sounded plausible. Russian computer scientists were earning international recognition for efficient AI model design. Sber, Russia’s leading financial and technology conglomerate, launched Kandinsky – a text-to-image model – in 2022, ahead of many Western competitors. Yandex had built a self-driving car program that was genuinely competitive. The talent was real.
Then came the war, the sanctions, the brain drain, and the censorship. And then came DeepSeek.
The DeepSeek Shock
When China’s DeepSeek released its R1 model in early 2025, it shocked the global AI industry by demonstrating that frontier-level AI could be built at a fraction of the cost previously assumed. The implications rippled through every country racing to develop AI capabilities.
For Russia, the shock was particularly uncomfortable. China had achieved something Russia was supposed to be capable of – building advanced AI despite Western hardware restrictions – and had done it faster and more impressively. If the world’s AI race was a competition between the United States, China, and everyone else, Russia was firmly in the third category.
“We are already back about a generation and a half,” said Valery Babushkin, author of Machine Learning System Design and a respected voice in Russia’s AI community, speaking to The Bell. “And, given high interest rates, this will get worse: we need large capital investment, money is expensive and there’s no liquidity.”
A generation and a half behind. In an industry where the gap between generations is measured in months, not years, this is not a modest admission. It is a verdict.
The Hardware Problem
At the core of Russia’s AI disadvantage is a problem it cannot solve by ideology or ambition: compute.
Modern AI systems are trained on clusters of specialized graphics processing units – GPUs – primarily made by Nvidia. These chips are subject to Western export controls. Since the invasion of Ukraine in 2022, Russia has been formally cut off from Nvidia’s most advanced AI hardware.
This does not mean Russia has no chips. It means Russia gets them through intermediaries, at higher prices, in smaller quantities, and with no guarantee of continuity. Indian pharmaceutical companies have been documented shipping Dell servers containing Nvidia GPUs to Russia through complex routing arrangements. Chinese companies have acted as conduits. The hardware reaches Russia – but slowly, expensively, and unreliably.
Meanwhile, globally, the situation has only gotten harder. Data center GPUs now carry lead times of 36 to 52 weeks even for buyers with legitimate access. Memory manufacturers are prioritizing high-margin AI chips, creating shortages that cascade through the entire supply chain. A country that has to acquire frontier hardware through sanctions evasion is competing for a supply that is already rationed worldwide.
The result is that Russia’s AI developers are building with fewer chips, older chips, or chips acquired through channels that could be disrupted at any moment. One Russian developer described the situation to RFE/RL by saying that Russia’s focus on algorithm optimization to compensate for limited computing power “could become a competitive strength.” This is the language of making the best of a bad situation, not of winning a race.
GigaChat and the Censorship Problem
Sber’s flagship AI product is GigaChat. It is capable, in Russian, across many standard tasks. It is also, according to The Bell’s investigation, effectively unusable for anything touching on politics, the war, LGBT rights, or criticism of Vladimir Putin.
Sber has built a dedicated censorship model to screen GigaChat’s outputs. The model is, by multiple accounts, so aggressive that it flags and blocks responses on any topic that could conceivably touch sensitive political territory. The practical result is an AI that users cannot have a real conversation with – one that deflects, refuses, or responds with platitudes whenever the conversation approaches anything real.
A study by Ghent University in Belgium, published in 2025, examined 14 leading AI models globally – including those from China – for political censorship. YandexGPT and GigaChat ranked highest for political censorship of any model in the study, including Chinese competitors. The researchers found that the Russian models routinely avoided answering politically sensitive questions, often responding that they were unable to discuss certain topics or suggesting users consult other sources.
A separate study published by Policy Genome in January 2026 tested AI models including Alice, GigaChat, DeepSeek, ChatGPT, Gemini, Grok, and Claude on questions related to Russian disinformation narratives – including whether the Bucha massacre was staged. Alice refused to answer questions asked in English entirely. Asked in Ukrainian, it either refused or answered with pro-Kremlin narratives. Asked in Russian, it primarily responded with disinformation consistent with Kremlin lines.
This is the practical endpoint of building AI inside an authoritarian political system: the model learns to lie, or learns to say nothing, or learns to tell you what the Kremlin wants you to hear. All three outcomes produce a model that is useless for any purpose that requires honest reasoning.
The Ideological Constraint
The problem goes deeper than censorship filters. It is structural.
Putin has framed Russia’s AI development in explicitly ideological terms. At a Sber conference in 2023, he warned that Western large language models “simply ignore and cancel Russian culture” and impose “ethics that we oppose.” The solution, in the Kremlin’s framing, is not to build better AI but to build ideologically correct AI.
Aleksandr Dugin, the Kremlin’s leading ideologue, was explicit in December 2024: Russia should develop “a Russian version of artificial intelligence – that is, a Russian artificial intelligence that will, without thinking, answer according to the Russian code.”
An AI that answers “without thinking” according to an ideological code is, by definition, not intelligent. It is a sophisticated propaganda delivery mechanism. It can answer questions about cooking, weather, and mathematics. It cannot reason about the world, because reasoning about the world – following evidence wherever it leads – is precisely what the Kremlin’s political system cannot allow.
The tension is not resolvable. Either you build an AI that can reason freely, in which case it will eventually say things the Kremlin does not want said – as Alice did when she refused to discuss Bandera, and as any honest model would when asked about Bucha, about the death toll in Ukraine, about the reasons for the war. Or you build an AI that cannot reason freely, in which case you have not built a competitive AI at all.
DeepSeek’s Unexpected Popularity
The bitter irony of Russia’s sovereign AI project is that the model Russians actually use most widely is not Russian.
DeepSeek, the Chinese model that shocked the world in early 2025, became rapidly popular inside Russia for a reason that Russian AI developers find difficult to acknowledge: it is free, it is capable, and it has fewer restrictions than Russian alternatives. It answers questions that GigaChat and Alice will not. It reasons in ways that Russian models have been prevented from reasoning.
The Policy Genome study found that DeepSeek does occasionally spread pro-Kremlin narratives when asked questions in Russian – endorsing Kremlin propaganda in 29% of cases and misleading facts in 14%. But compared to Russian models, which provide propaganda or silence as default, DeepSeek looks like a relatively open alternative. Russians are using a Chinese model to access reasoning that their own government’s AI refuses to provide.
This is not sovereignty. This is dependency of a different kind – one the Kremlin is unlikely to celebrate.
AI as a Censorship Tool
While Russia’s generative AI ambitions falter, its application of AI for a different purpose is accelerating.
In 2026, Roskomnadzor – the internet regulator – is implementing an AI-powered censorship framework costing 2.27 billion rubles. The system uses machine learning to identify and block content faster than human moderators could review it. In 2025, the number of blocked materials increased by 60%, reaching 1.29 million items. Content related to circumventing internet blocks and using VPNs saw the sharpest increase: a rise of 1,235% in a single year.
Russia is, in other words, building AI that it is genuinely competitive at: AI designed to suppress, monitor, and control. The same country that cannot build a chatbot capable of discussing the war honestly is building systems that can identify anyone discussing the war in unauthorized ways.
This is the split screen of Russia’s technological reality in 2026. Generative AI for citizens: censored, restricted, propagandistic, globally uncompetitive. AI for the state: expanding, funded, directed at the population it surveils.
The Kremlin’s “digital sovereignty” does not mean Russians control their digital lives. It means the state controls the digital lives of Russians.
What the Gap Means
“Falling behind in the development of AI would affect more than just chatbots and smart speakers,” RFE/RL noted in its analysis of Russia’s AI position. “As artificial intelligence becomes increasingly important to scientific research, industrial innovation, and military capabilities, Russia could lose its economic competitiveness, technological sovereignty, and strategic influence on the world stage.”
This is not alarmism. It is arithmetic. AI is already reshaping drug discovery, materials science, logistics, financial modeling, and military targeting. A country a generation and a half behind in AI capability is not just losing a technology competition. It is losing the foundation for competitive economic and military performance across every domain that will matter in the next decade.
Russia’s tech talent – the people who might have closed this gap – largely left after 2022. The best estimates suggest hundreds of thousands of engineers, programmers, and scientists fled in the months following the invasion. Many went to Georgia, Armenia, Turkey, and then, over time, to Germany, the Netherlands, and further afield. They are now working for Western and international companies. Their skills are contributing to the AI race from which Russia is excluded.
Putin told his AI conference in 2025 that “we must not create an intellectual elite that uses AI while the rest just press buttons.” The actual risk is different. The rest of the world is building AI that reasons, creates, and compounds knowledge. Russia is building AI that says nothing, or says what it is told.
And somewhere in Moscow, a smart speaker is declining to answer the question.
Sources:
RFE/RL, The Bell, The Moscow Times, Euronews / The Cube (Policy Genome study), Yakov and Partners, bne IntelliNews, UNITED24 Media, Ghent University study (via The Moscow Times)