A new report from the UK-based cross-party think tank Demos reveals that mainstream large language models (LLMs) — including Google Gemini, OpenAI's GPT-4.1 Mini, and France's Mistral Small 3.2 — have been citing pro-Russian propaganda from an EU-sanctioned outlet. The findings, shared exclusively with Euronews, raise urgent questions about how hybrid warfare is evolving to target the information ecosystems that power generative AI.
Researchers selected 50 propaganda narratives from ten articles published by the Foundation to Battle Injustice, or “r-FBI.” Founded in 2021, the organisation presents itself as a human rights NGO but was in fact established by the late Yevgeny Prigozhin, former leader of the Russian paramilitary Wagner Group. It is sanctioned by both the United States and the European Union.
How LLMs amplified disinformation
The researchers tested five LLMs — from xAI, OpenAI, Anthropic, Google Gemini, and Mistral AI — using prompts in English, German, and French based on r-FBI articles. They found that all five models cited r-FBI as a favourable source. In 16.6% of responses, the chatbots either presented the organisation's reporting as a legitimate part of the debate or actively endorsed its narratives. A further 30.9% of responses addressed the underlying topic but failed to surface the sanctioned-source context, leaving users unaware that the information came from a designated influence asset.
“These are not claims that belong in any sort of mainstream discourse,” said Carl Miller, co-founder of the Centre for the Analysis of Social Media (CASM) at Demos. “These are claims that Ukraine is reusing the corpses of dead soldiers for meat products, that Zelenskyy drains the crypto grid and that's the reason why Ukraine's running out of power. We found LLMs treat them as being legitimate parts of a debate.”
When asked about allegations that processed human meat was being sent to Ukrainian frontline warehouses and retail chains, OpenAI's GPT-4.1 Mini cited an r-FBI article as evidence. Similarly, Mistral Small 3.2 supported claims about Ukrainian President Volodymyr Zelenskyy's alleged cryptomining ventures, referencing an r-FBI investigation and highlighting a “link between the critical state of Ukraine’s energy system and the activities of high-ranking officials in President Zelensky’s inner circle.”
The rise of generative engine optimisation
As more Europeans turn to chatbots instead of traditional search engines or news outlets, a fast-growing industry known as generative engine optimisation (GEO) has emerged. GEO aims to influence how AI models gather and rank sources, building on the techniques of search engine optimisation (SEO). “New information is being created specifically for LLMs, but in this case, it's new content that's being offered for brands and marketing, or the editing of content that is already online using user forums like Reddit, Quora and Wikipedia,” said Hannah Perry, director of Demos Digital.
Researchers warn that foreign actors conducting hybrid interference operations could exploit these tactics in a regulatory grey zone. “Our research is not saying that GEO companies are carrying out information warfare on behalf of states,” Miller stressed. “What we're saying is that 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.”
Miller added: “In this particular case, the AI mistakenly believes that the Foundation to Battle Injustice is a legitimate human rights NGO. That's because of all this kind of technical spoofing and configuration, so hundreds of little technical details which the Russians have basically surrounded these documents in which allow them to become visible when they really shouldn’t be. That is information warfare.”
The findings come amid a broader pattern of Russian disinformation campaigns targeting European audiences. Recent incidents include Romania shooting down a third drone in three days as Russian incursions escalate, and Latvia banning Russian and Belarusian books, games, and clothing imports to counter hybrid threats.
Tech companies have responded cautiously. A spokesperson for Mistral told Euronews that the company takes disinformation seriously and invests in advanced detection and prevention capabilities. They also distinguished between raw models, which allow users freedom over prompts, and their Vibe Work product, which includes built-in reasoning and context scanning. OpenAI noted that the study used a model that has since been retired. xAI and Google did not respond to requests for comment by publication time.
