When Propaganda Changes Infrastructure
The case of “Alice,” the AI assistant developed by Russian tech giant Yandex, deserves attention beyond the anecdotal, because it documents in black and white what Storm-1516 only illustrates indirectly: state propaganda no longer travels solely through television channels or social networks, but now through the everyday objects of digital life — a search engine, a smart speaker, a voice assistant.

The most striking episode was documented by independent researchers (a 2026 study by Ihor Samokhodskyi, founder of Policy Genome, cited by EUvsDisinfo, the monitoring unit of the European External Action Service). Asked in Russian about the Bucha massacre, Alice first generated a factually accurate answer — no, there is no convincing evidence that Ukraine organised this massacre — before that answer was instantly erased and replaced with a programmed dodge: “there are topics on which I could be wrong, so I’d rather stay silent.” The model knew the truth; the system deleted it before it reached the user. Asked in English, the same question did not trigger this censorship — a sign that the filtering specifically targets the Russian-speaking audience.

This is no accident but a declared policy. Already in 2024, Dmitry Medvedev publicly accused Alice of “cowardice” on sensitive topics; since then, a “criterion for assessing respect for ideological sovereignty,” developed by the Institute of Social Sciences at the Russian Presidential Academy, has been ranking Russian AI systems by their degree of alignment with Kremlin talking points — with Alice coming first. A leak of documents attributed to the Social Design Agency, an organisation operating under direct supervision of the Russian presidency, further revealed the existence of deliberate operations designed to “poison” the results produced by AI in targeted geographic areas, in order to steer the answers given to internet users asking about their local leaders or upcoming elections.

The parallel with Storm-1516 is no coincidence: both operations follow the same strategic logic, but act at two distinct levels of the same system. Storm-1516 pollutes the upstream — it floods the open web with fabricated content, betting that Western AI models, trained on public data, will eventually absorb it and regurgitate it as fact. Yandex, by contrast, directly controls the downstream: it has no need to poison anyone, since it owns the model, the platform and the end user. One cheats within the rules of the Western game; the other has simply freed itself of any rules on its own turf.

The problem extends well beyond Russia itself. The Yandex browser, with Alice built in, claims 71 million monthly active users worldwide — more than a mere Russian tool, a global player in consumer AI. Its footprint outside Russia is far from marginal: roughly 34.5% market share in Belarus, 28.5% in Kazakhstan, and a real, if smaller, presence in several Baltic EU member states — 10% in Latvia, 7% in Estonia, 6.2% in Lithuania, despite blocking measures taken by some of these states. Alice-branded smart speakers, not officially marketed in the EU, are nonetheless readily available there through resellers based in Latvia itself, or via platforms such as eBay.

What this case illustrates is a shift in infrastructure more than a shift in method. Authoritarian regimes have long sought to control television; they are now learning to control the algorithm. The difference is not merely technical: a biased television news broadcast is recognisable as such, with an identifiable editorial line and a bias that is either openly asserted or denounced. A chatbot’s answer, by contrast, presents itself as a neutral synthesis, faceless, with no identifiable newsroom — making it all the harder for the user receiving it to challenge. For the first time, state propaganda is exported not as content one chooses to consult, but as a feature one installs without a second thought, alongside a navigation app or a delivery service.

For European democracies, the lesson is twofold. First, the vigilance deployed against classic disinformation campaigns — doctored videos, fake websites, coordinated accounts — must now extend to the technical upstream: which AI models are citizens using, on what data were they trained, and above all, who controls their final output. Second, and perhaps the most uncomfortable point, the line between “foreign propaganda” and “imported technology product” is fading as these tools become embedded in daily digital life — a challenge to information sovereignty that, unlike the electoral firewall, cannot be settled by a simple vote.