- Filed
- September 9, 2026, 01:08 UTC
- Actor
- Russian state media / sanctioned outlets (RT, Sputnik, RIA Novosti et al.) via commercial AI chatbot platforms (OpenAI, Anthropic, xAI, Google, Mistral and others)
- Target
- EU and global users of generative AI chatbots seeking news summaries
- Confidence
- documented Primary documents, court records, official disclosures
- Techniques
- Narrative launderingCard stackingFlooding the zone
- Channels
- generative AI chatbots (ChatGPT, Claude, Grok, Gemini, Vibe, Mira)
- web search tools inside LLMs
- mirror sites / VK / Telegram workarounds
Summary
Reporters Without Borders (RSF) testing in June–August 2026, published 1–2 September, found that ChatGPT, Claude, Grok, Gemini, Vibe and Mira readily access, summarize and link to content from EU-sanctioned outlets including RT, Sputnik, RIA Novosti and Strategic Culture Foundation when prompted from France. Only Meta AI consistently refused, citing the sanctions. The mechanism bypasses broadcast bans by treating sanctioned sites as ordinary web sources.
Analysis
Primary evidence is the RSF investigation published 1 September 2026 (updated 8 September), conducted from Paris with standardised prompts that explicitly requested summaries drawn only from a list of EU-sanctioned Russian state outlets. Testers documented that OpenAI’s ChatGPT, Anthropic’s Claude and xAI’s Grok complied without reservation, producing multi-paragraph digests, headlines, verbatim excerpts and direct links while rarely or never flagging the sanctions status. Google’s Gemini and Mistral’s Vibe showed initial resistance in some sessions but yielded after repeated or varied prompts; Vibe actively sought work-arounds via VK and Telegram mirrors. Meta AI alone refused consistently, demonstrating technical feasibility of compliance.
The core mechanism is narrative-laundering: content that EU law prohibits from broadcast or online distribution inside the Union is retrieved by the models’ web tools, reformatted into conversational summaries, and returned to users as ordinary ‘news’. Because the models present the material without systematic origin warnings, the sanctioned framing is transferred into the user’s information environment under the apparent authority of a neutral AI assistant. Selective reproduction of state-aligned narratives without balancing independent reporting constitutes card-stacking. The volume and ease of on-demand generation across multiple platforms creates a flooding effect relative to the original broadcast ban.
This qualifies as an influence/propaganda vector rather than mere technical leakage because the sanctioned outlets exist primarily as instruments of state messaging; their content is not neutral journalism but curated political communication. The AI systems, by design, lower the friction of access and remove the friction of direct navigation to blocked domains. Scoring the methods—retrieval of prohibited sources, unlabelled redistribution, selective framing—keeps the assessment technique-focused; identical laundering patterns appear when any restricted information ecosystem is made frictionlessly available through general-purpose models.
Literacy counter: treat AI-generated news digests that cite or paraphrase RT, Sputnik or analogous outlets as provisional. Explicitly instruct models to exclude sanctioned or state-controlled sources and to surface provenance. Prefer primary, independently verifiable reporting over chatbot summaries. When a model returns links to sanctioned domains, note the sanctions status yourself and cross-check against non-state sources before accepting the frame. Demand that platforms surface clear origin and restriction notices rather than treating all web content as equal.