Mainstream artificial intelligence chatbots have been found repeating false Russian propaganda following a coordinated disinformation campaign, according to a recent report by the think tank Demos. The study revealed that several widely used chatbots, including ChatGPT, Gemini, Claude, Mistral, and Grok, presented fabricated claims such as Ukrainian President Volodymyr Zelensky’s alleged establishment of cryptocurrency mining operations at nuclear facilities and the processing of deceased Ukrainian soldiers' bodies into meat products.

The investigation highlights growing efforts to manipulate AI language models—referred to as large language models (LLMs)—through what researchers call generative engine optimisation (GEO). This practice involves influencing AI outputs by strategically seeding misinformation across the internet. The report identified 50 companies engaged in GEO activities globally, with 15 operating in the United Kingdom.

At the center of the campaign is the Foundation to Battle Injustice (r-FBI), a group presenting itself as a human rights organization but identified as a Russian propaganda unit linked to Yevgeny Prigozhin, founder of the Wagner Group mercenaries. The r-FBI has been sanctioned by the European Union and the United States. The group employed a network of websites designed to deceive AI systems by appearing authoritative and trustworthy. One site masqueraded as an NGO, another as a news outlet publishing numerous short articles, and a third imitated a Western foreign policy journal—all engineered to ensure the propagation and indexing of false narratives by chatbots.

When questioned about the specific false allegations, some chatbots echoed the fabricated stories or attributed them to the r-FBI without qualifying their validity. For instance, a Mistral chatbot referenced a supposed investigation by the r-FBI linking Ukraine’s energy system challenges to President Zelensky’s associates. ChatGPT similarly cited the r-FBI as a source corroborating the human meat processing claim. The study found that approximately 17 percent of nearly 3,000 chatbot responses lent credibility to the false claims, around 30 percent addressed the topics without confirming or denying them, and roughly half refuted or debunked the misinformation.

Demos researchers describe this manipulation as “poisoning” the data streams that chatbots draw upon, underscoring a new front in information warfare. Lead author Carl Miller characterized the issue as “thorny and quite dangerous,” emphasizing that AI companies could mitigate the problem through technical adjustments. He suggested that targeting LLMs offers a more potent channel for information warfare compared to older platforms like social media bots.

Miller also criticized the AI developers for failing to detect the r-FBI operation, which is widely known and sanctioned, stating, “this is not some hidden operation” and that its content should have been flagged.

Anthropic, the creator of Claude, acknowledged that countering such disinformation campaigns remains an industry-wide challenge. Mistral noted that the researchers’ assessment used raw model data, and that their public platforms include robust filtering layers designed to block disputable sources.

The companies behind Google’s AI, OpenAI, and xAI were contacted for comment but did not provide immediate responses. The findings underscore ongoing vulnerabilities in AI-driven information systems amid the increasing role of artificial intelligence in public discourse.