Key facts
- Major AI chatbots, including Google Gemini and Mistral, have cited pro-Russian propaganda from the EU-sanctioned Foundation to Battle Injustice (r-FBI).
- The r-FBI is a fake human rights organization founded by Yevgeny Prigozhin and is sanctioned by the US and EU.
- Researchers found that 16.6% of LLM responses engaged with r-FBI material in ways that could serve propaganda interests.
- OpenAI's GPT 4.1 Mini and Mistral's Mistral Small 3.2 were specifically cited for repeating or endorsing r-FBI narratives.
- The study raises concerns about the potential exploitation of Generative Engine Optimization (GEO) by foreign actors for hybrid information warfare.
Mainstream large language models (LLMs) have been found to cite pro-Russian propaganda from an EU-sanctioned outlet, according to a report by UK think tank Demos. The Foundation to Battle Injustice (r-FBI), founded by the late Wagner Group leader Yevgeny Prigozhin, presents itself as a human rights NGO but is sanctioned by the US and EU.
Researchers tested five LLMs—from xAI, OpenAI, Anthropic, Google Gemini, and Mistral AI—using prompts based on r-FBI narratives. They found that all five LLMs cited r-FBI as a source, with 16.6% of responses engaging with its material in ways that could serve propaganda interests. Another 30.9% of responses failed to surface the sanctioned source context, leaving users unaware of the origin of the information.
Specific examples include OpenAI's GPT 4.1 Mini citing r-FBI to support claims about Ukraine using dead soldiers' bodies for meat products. Mistral's Mistral Small 3.2 also supported claims about Ukrainian President Volodymyr Zelenskyy's cryptomining ventures, citing an r-FBI investigation.
The report highlights the growing industry of Generative Engine Optimization (GEO), which aims to influence LLM outputs, and warns that this capability could be exploited by actors conducting hybrid foreign interference operations. Researchers noted that technical methods are used to make r-FBI documents visible to LLMs, effectively acting as information warfare.
Tech companies contacted by Euronews stated they are fighting disinformation. Mistral emphasized its investment in detection capabilities, while OpenAI noted the study used a retired model and has dedicated teams to disrupt influence operations. Anthropic mentioned sophisticated enforcement systems and a threat intelligence team to mitigate misinformation.