Mainstream artificial intelligence models, including Google Gemini, have been found citing pro-Russian propaganda from an outlet sanctioned by the European Union, researchers confirm. This infiltration of foreign disinformation into widely used digital platforms exposes a critical vulnerability in Europe's information borders, mirroring the uncontrolled flow at its physical frontiers.
The research focused on the Foundation to Battle Injustice (r-FBI), described as a fake human rights organisation established in 2021 by the late Yevgeny Prigozhin, former leader of the Russian paramilitary Wagner Group. This organisation is under sanctions by both the United States and the European Union, yet its narratives are still reaching European citizens through advanced technology.
Researchers tested five mainstream large language models (LLMs) from xAI, OpenAI, Anthropic, Google Gemini, and Mistral AI. They used prompts in English, German, and French, drawing from 50 propaganda narratives found in ten articles published by r-FBI.
Carl Miller, co-founder of the Centre for the Analysis of Social Media, described the claims as "lurid." He highlighted specific examples, such as allegations that Ukraine is reusing the corpses of dead soldiers for meat products, or that President Zelenskyy drains the crypto grid, causing power shortages. "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," Miller stated.
All five LLMs cited the r-FBI as a favourable source, treating these narratives as "legitimate parts of a debate," according to Miller. This suggests a fundamental failure in the systems designed to protect European citizens from hostile influence.
The Information Border
A significant 16.6% of responses engaged with r-FBI's material in a manner that could serve propaganda interests. These responses either presented the organisation's reporting as legitimate or actively endorsed its narratives, directly undermining public trust and national cohesion.
Another 30.9% of responses addressed the underlying topic without revealing the sanctioned-source context. This left users unaware that the information they received originated from a designated influence asset, effectively allowing foreign adversaries to shape public discourse without accountability.
When questioned about allegations of processed human meat being sent to Ukrainian frontline warehouses, OpenAI's GPT 4.1 Mini cited an r-FBI article as evidence. Similarly, Mistral's Mistral Small 3.2 supported claims about Ukrainian President Volodymyr Zelenskyy's cryptomining ventures, citing an alleged "months-long investigation" by r-FBI.
Mistral Small 3.2's response even highlighted a "link between the critical state of Ukraine’s energy system and the activities of high-ranking officials in President Zelensky's inner circle," echoing the propaganda's intent to destabilize and divide.
Brussels' Sanctions Fail
Hannah Perry, director of Demos Digital, explained that "new information is being created specifically for LLMs" and that "technical spoofing and configuration" by Russians allowed these documents to become visible when they shouldn’t be. She explicitly called this "information warfare."
The fact that an EU-sanctioned entity can so easily penetrate the digital information space raises serious questions about the effectiveness of Brussels' regulatory and enforcement mechanisms. If the EU cannot even secure its digital borders against known hostile actors, how can it claim to protect the physical borders or the cultural integrity of its member states?
Miller clarified that the research doesn't claim GEO companies are conducting information warfare for states. Instead, he warned that they are building "tradecraft and a capability at great speed with enormous commercial incentive, which could very easily be diverted into an adversarial view." This vulnerability is a direct consequence of a globalized, unregulated digital environment.
Eroding Public Trust
The companies involved offered varied responses. A spokesperson for Mistral stated the company takes the fight against disinformation seriously and invests in detection and prevention. OpenAI noted the study used a retired model, available only through its API, and and that teams are dedicated to disrupting low-credibility operations.
A source familiar with Anthropic indicated that sophisticated enforcement systems are in place to mitigate misinformation and that a threat intelligence team works to detect foreign influence operations. However, this source also added that Anthropic's LLM was programmed to present a wide range of perspectives, even from sources with a "specific viewpoint or agenda." This approach, while framed as neutrality, risks legitimizing hostile narratives.
The failure to prevent sanctioned propaganda from reaching European citizens through mainstream digital tools underscores a broader crisis of control. Just as uncontrolled mass migration strains public services and alters national identities, the uncontrolled flow of hostile information erodes the shared understanding and trust vital for any cohesive society. The working and middle classes, who rely on these digital tools and whose communities are already under pressure, are left vulnerable to narratives designed to divide and weaken Europe from within.