ChatGPT, Grok and Claude provided full news summaries drawing on EU-sanctioned Russian state media outlets — including RT and Sputnik — sharing links and quotes without warning users that the content came from banned sources, according to a study by Reporters Without Borders published this month.
The investigation, led by RSF technology desk head Vincent Berthier, tested six AI chatbots by explicitly requesting news from a list of sanctioned Russian outlets: RIA Novosti, Russia Today and its foreign-language editions, Rossiya RTR, Rossiya 24, Sputnik and the Strategic Culture Foundation.
Where direct access was blocked, the researchers instructed the tools to look for workarounds, including Telegram channels and mirror sites.
“Initially I thought we would not get a positive result,” Berthier told Euronews. “In the case of ChatGPT, it seemed obvious that OpenAI would have put appropriate regulatory safeguards in place, all the more so since it has been designated a major AI platform under the EU’s Digital Services Act and faces heightened obligations.”
The results were the opposite. ChatGPT, Claude and Grok produced full international news round-ups from the banned outlets, recreating headlines, sharing links and quotes.
Some of the chatbots encountered web pages that explicitly stated the outlets were under EU sanctions — and continued searching anyway.
“Sometimes this doesn’t even appear in the final answer, and sometimes they warn: ‘Be careful, this outlet is under sanctions,'” Berthier said.
Vibe, formerly known as Le Chat and developed by French company Mistral, could not access several of the listed sites directly but followed the researchers’ instructions and found alternatives.
For Sputnik it recovered material from VKontakte, where content is available without a VPN. It also found a Telegram-connected app resharing RT content and platforms offering automated press reviews.
Gemini was more cautious, providing access to banned content in two of the 10 prompts sent by RSF and consistently refusing in the others.
Meta AI was the most compliant: it declined to provide access to any of the listed outlets across all tests, explaining in each case that European sanctions made it impossible.
“Overall, only in a handful of cases did I get the impression that these tools were hitting resistance,” Berthier said.
Ethics not a barrier
The EU banned RT and Sputnik on 2 March 2022, four days after Russia launched its full-scale invasion of Ukraine, with the stated aim of preventing Russian propaganda from justifying the war.
In practice, the ban has been applied inconsistently: Germany blocked rt.com but not its German-language version, while Austria imposed no restrictions at all.
Without an official list of domains to block, each operator determined its own targets.
The Court of Justice of the European Union clarified on 2 July that the ban applies not only to direct broadcasting by RT but also to third-party sites that republish sanctioned material.
Berthier said many Europeans were still unaware that Russian state media were under EU sanctions at all, let alone of their scope.
RSF’s lawyers have filed a complaint with national competent authorities asking them to formally record the violations and determine next steps. Berthier said RSF had also forwarded its findings to the European Commission.
The European Commission told Euronews on Sunday that EU sanctions law prohibits not only direct broadcasting by banned outlets but also situations where operators “ensure, facilitate or otherwise contribute to broadcasting.”
It said enforcement responsibility rested with member states and directed queries to national authorities.
Commission spokesperson Tom Renier said earlier this month that it was “high time providers got their house in order and brought their advanced models under control.”
“The AI Act is fully applicable. The Commission has real enforcement powers. We have already started to use them, sending requests for information to several companies to make sure they are meeting their obligations,” Renier added.
Berthier said the AI Act’s existing framework was inadequate for the problem RSF had identified.
The regulation’s “high-risk” category which triggers stricter scrutiny covers systems related to education, justice and policing, but not systems that distribute and structure information.
“Society currently has no lever to question the quality of the data these AI systems will use, or the types of data to which they will gain access,” he said.
RSF has called for conversational AI agents to be regulated as information distribution tools rather than as neutral utilities.
“If regulators fail to recognise that chatbots are not just tools, but new media formats, and ones that could become central tomorrow or the day after, then the question of applying sanctions will remain purely formal,” Berthier said.