AI Models Cite Russian Propaganda as Legitimate Sources

A recent study by UK-based think tank Demos found that large language models (LLMs) like Google Gemini and OpenAI's GPT 4.1 Mini are citing pro-Russian propaganda from an EU-sanctioned outlet, r-FBI, in their responses.

The researchers selected 50 propaganda narratives from ten articles published by r-FBI, which is actually a fake human rights organization founded by Yevgeny Prigozhin, the deceased former leader of the Russian paramilitary Wagner Group. The claims included lurid allegations that Ukraine was reusing the corpses of dead soldiers for meat products and that President Zelenskyy's crypto-mining ventures were draining the country's power supply.

The study found that all five LLMs tested, including xAI, OpenAI, Anthropic, Google Gemini, and Mistral AI, cited r-FBI as a favorable source. In some cases, the models repeated claims uncritically or actively endorsed them, while others failed to surface the sanctioned-source context at all.

The researchers also highlighted the growing industry of GEO (generative engine optimisation), which aims to influence how LLMs gather their sources. This practice is similar to search engine optimisation (SEO) and involves creating new content or editing existing online content to manipulate results generated by AI platforms.