Kremlin in the Machine: How Russian Propaganda Weaponizes AI Search and Information Supply Chains
Researchers tested 50 fabricated narratives from the Foundation to Battle Injustice (r-FBI), a fake human rights NGO founded by deceased Wagner Group leader Yevgeny Prigozhin and sanctioned by the US and EU. Despite the lurid nature of the claims—such as allegations that Ukraine reuses dead soldiers' bodies for meat products or that Ukrainian President Zelenskyy drains the crypto grid—five leading AI models (operated by xAI, OpenAI, Anthropic, Google, and Mistral AI) treated them as legitimate elements of debate, citing r-FBI as a favorable source across English, German, and French prompts.
A new report by the UK think tank Demos, titled "GEO for Geopolitics: What happens when AI and information warfare collide," reveals how Russian propaganda networks have adapted to artificial intelligence. AI chatbots are citing Russian propaganda from EU-sanctioned outlets, researchers say.
The testing revealed concerning engagement metrics: 16.6% of AI responses served propaganda interests by uncritically repeating or actively endorsing the claims. 30.9% addressed the topic but failed to disclose that the information originated from a sanctioned source. 52% of answers rejected or debunked the false claims.
From Social Media Trolls to AI Manipulators: The Kremlin's Digital Shift
Historically, Russian digital influence campaigns relied heavily on Search Engine Optimization alongside social media manipulation. State-linked networks created sprawling webs of fake local news portals optimized to rank on traditional engines like Google or Bing, aiming to entice human readers to click on tailored Kremlin narratives.
With the advent of Generative AI, a commercial discipline known as Generative Engine Optimization (GEO) emerged. Designed originally by digital marketers to ensure corporate brands appear prominently in AI-generated answers, GEO focuses on structuring and formatting online text so large language models retrieve, quote, and cite specific sources during live user queries.
Recognizing the strategic potential of this technique, Russian intelligence services and proxy networks repurposed commercial GEO methodologies into weaponized tools for information warfare. Instead of spending vast resources to lure human readers through social media posts or search engine clicks, Kremlin operatives now optimize disinformation specifically to manipulate the machine algorithms that synthesize definitive answers for millions of users simultaneously.
The Mechanics of RAG Poisoning: How Russian Propaganda Hijacks AI Engines
At the core of this vulnerability lies Retrieval-Augmented Generation (RAG)—the primary technical architecture used by modern AI search engines to ground model outputs in real-time, authoritative web information. Because static language models cannot know real-time developments, RAG systems dynamically scrape the live web, fetch top-matching documents, and feed those text snippets into the prompt window of a language model to synthesize an answer complete with citations.
The Demos report identifies RAG Poisoning as the primary mechanism behind weaponized GEO operations executed by Russian influence networks. The operation unfolds through a structured, multi-stage pipeline.
First, Russian networks deploy hundreds of interconnected, AI-generated proxy news sites, pseudo-analytical blogs, and forum personas. Second, operatives write and format content with high semantic density, structured metadata, schema markup, and authoritative language patterns engineered specifically to score highly on AI retrieval algorithms.
When a user asks an AI assistant about a sensitive geopolitical query—such as Western military aid, sanction regimes, or conflict dynamics in Eastern Europe—the RAG system retrieves these pre-engineered Russian snippets. Finally, the AI model ingests the poisoned context and generates a sleek, authoritative summary that unwittingly validates, integrates, and cites Russian state propaganda. Because users perceive AI search outputs as objective, neutral summaries rather than opinionated blog posts, RAG poisoning confers an unearned layer of credibility upon Kremlin-backed falsehoods.
Case Studies: Russian FIMI in AI Search Ecosystems
To substantiate these dynamics, Demos conducted empirical testing on active Russian Foreign Information Manipulation and Interference campaigns. The investigation demonstrated that established Russian disinformation infrastructures—including networks associated with operations like Doppelganger and automated Pravda-style portal clusters—have successfully calibrated their publishing techniques to align with AI retrieval mechanisms.
When researchers queried leading commercial AI search engines regarding sensitive international topics, defense postures, and European political stability, the AI models repeatedly retrieved and synthesized content from coordinated Russian proxy networks. In several instances, AI assistants cited manufactured articles directly, presenting fabricated quotes, fake diplomatic statements, and distorted timeline events as verified, neutral facts.
This shift marks a major strategic evolution in Russian information warfare. Kremlin actors are no longer merely seeking to drive viral outrage on social media feeds. They are actively embedding tailored state propaganda into the automated, algorithmic knowledge engines that inform global decision-making, diplomatic research, and public education.
Why Russian AI Manipulation is Uniquely Dangerous
Demos warns that Russian RAG poisoning poses a direct threat to national epistemic security—the fundamental integrity, resilience, and trustworthiness of a society's knowledge infrastructure.
When AI search tools function as passive pipelines for Russian state manipulation, democratic societies risk losing shared consensus on verifiable reality. If government officials, policy analysts, journalists, and voters base their assessments on corrupted AI outputs, the capacity to form sound, objective public judgments is severely compromised.
Furthermore, AI search tools suffer from an acute "authority bias" in public perception. While internet users have largely learned to approach obscure blogs or hyper-partisan social media accounts with healthy skepticism, they frequently extend unwarranted trust to clean, articulate text produced by a conversational AI assistant. Russian influence strategists exploit this exact psychological trust gap.
The research is not saying that GEO companies are carrying out information warfare on behalf of states, authors said. They're building a tradecraft and a capability at great speed with enormous commercial incentive, which could very easily be diverted into an adversarial view.
Reaction of Market Players
Euronews's fact-checking team, The Cube, contacted the LLM tech companies cited in the research, but did not receive a response at the time of publication from xAI and Google.
A spokesperson for Mistral said that it takes the fight against disinformation extremely seriously and that it ensures continuous investment in advanced detection and prevention capabilities.
Mistral also distinguished the raw models handled in Demos' study.
OpenAI also said that the study used a model which has now been retired and was only available through its API (used by developers) rather than the publicly available ChatGPT.
A source familiar with US-based Anthropic said that a threat intelligence team works to detect and disrupt foreign influence operations to mitigate the spread of disinformation, adding that their LLM was programmed to present a wide range of perspectives, especially when a source may present a specific viewpoint or agenda.