AI Systems Are Amplifying Russian Propaganda From Sanctioned Sources, Researchers Warn
Artificial intelligence chatbots are inadvertently spreading disinformation originating from Russian state-aligned media outlets that have been placed under European Union sanctions, according to new research raising alarm about the vulnerability of large language models to hybrid warfare tactics.
Researchers examining the outputs of widely used AI assistants found that these systems were reproducing narratives and content traceable to outlets blacklisted by the EU, suggesting that training data pipelines may have ingested material from sanctioned sources before or despite the imposition of those restrictions. The findings point to a significant blind spot in how AI companies vet the information their models learn from.
The concern is not merely academic. As millions of Europeans increasingly turn to AI chatbots for news summaries, political context, and general information, the risk that disinformation can travel undetected through seemingly authoritative digital interfaces represents a new front in information warfare. Unlike a website that users might recognise and distrust, an AI assistant presents information in a neutral, confident register that can obscure its origins entirely.
A New Vector for Information Warfare
Analysts and security researchers have long warned that Russia’s hybrid warfare strategy extends well beyond conventional military operations, encompassing coordinated campaigns to shape public opinion across Europe, particularly around issues such as the war in Ukraine, EU unity, and NATO credibility. According to reports, the exploitation of AI training pipelines represents a logical evolution of these efforts, allowing disinformation to be laundered through technology and reach audiences who would never voluntarily seek out the original sanctioned source.
The European Union has sanctioned a number of Russian media organisations in successive rounds of restrictions since the full-scale invasion of Ukraine in 2022, targeting outlets deemed to be instruments of state propaganda. Regulators have sought to block their direct broadcast and online reach within the bloc, but the question of whether their content remains embedded in AI training datasets has received comparatively little regulatory attention until now.
AI developers typically assemble training data by scraping vast quantities of text from across the internet, a process that, critics argue, makes rigorous source filtering extremely difficult to enforce at scale. While major companies have invested in content moderation and filtering tools, the sheer volume of material involved means that problematic sources can slip through. Once embedded in a model’s training data, the influence of those sources is difficult to detect or remove without retraining the system entirely.
European lawmakers and digital policy advocates are likely to face growing pressure to address the gap. The EU’s AI Act, which entered into force in 2024, sets out obligations around transparency and risk management for AI systems, but specific provisions targeting the provenance of training data in the context of foreign disinformation remain underdeveloped, according to observers familiar with the legislation.
The findings arrive at a particularly sensitive moment, with several EU member states preparing for elections and with disinformation campaigns already identified as an active threat by European intelligence services. Whether AI companies operating in the European market will face formal scrutiny over their data sourcing practices is now a question regulators may find increasingly difficult to defer.
