Ask a chatbot for the news, and it might quietly hand you Kremlin propaganda. A new investigation found that ChatGPT, Claude and Grok all serve up content from Russian outlets banned in the EU, without ever mentioning the ban. Only one major AI assistant refused to play along.

The European External Action Service (EEAS), headed by High Representative Kaja Kallas, is closely monitoring the issue, a Commission spokesperson told EU Perspectives. It “is closely monitoring developments in the field of AI and the role it plays both in the dissemination of illegal or false information and in potentially countering such content,” the spokesperson said.

The service is also “in contact with AI providers as part of its ongoing efforts to monitor these developments”. This is meant to “implement appropriate countermeasures to detect FIMI activity”, or foreign information manipulation and interference.

Bypassing sanctions

Accused of being Kremlin “disinformation” tools, outlets such as Sputnik and RT were banned from broadcasting in the EU after Russia’s invasion of Ukraine in February 2022. Yet according to Reporters Without Borders (RSF), when asked directly, ChatGPT, Claude and Grok, the chatbot built into the X platform, all agreed to circumvent the sanctions and reproduce the outlets’ content.

The chatbots “provided a comprehensive press review of international news, reporting headlines and sharing links and text statements, without systematically indicating that the outlets in question were subject to European sanctions”, the NGO explained. RSF ran the tests in June and August.

Contacted by RSF, OpenAI said in a statement that it takes the findings seriously and intends to examine the evidence more closely. After some resistance and repeated requests, Mistral’s Vibe (formerly Le Chat) and Google’s Gemini also responded to the test questions. Only Meta AI refused to comply, citing the EU sanctions, which RSF says “proves that it is possible for a chatbot to be compliant”.

Under these conditions, the sanctions adopted by the EU are failing to achieve their objective. — Vincent Berthier, head of technology and journalism at RSF

“Under these conditions, the sanctions adopted by the EU are failing to achieve their objective”, said Vincent Berthier, head of technology and journalism at RSF. The NGO is calling on Brussels to launch an investigation to determine “the extent to which OpenAI is failing to meet its obligation to prevent systemic disinformation risks on its ChatGPT service”, and also hopes to open infringement proceedings against the other chatbots.

A regulatory grey zone

Chatbots can be highly effective tools for circumventing EU sanctions by exploiting vulnerabilities in how they source information online. “This investigation demonstrates, above all, the need for ex ante regulation of AI-based chatbots: otherwise, ex post sanctions will remain more symbolic than effective”, Berthier explained.

The Commission has confirmed that ChatGPT falls within the scope of the Digital Services Act (DSA). RSF is calling on the executive to launch an investigation immediately.

Yet, as a Commission spokesperson pointed out, “the DSA does not regulate specific content, but rather platforms. In these risk assessments, providers must consider, among other things, how their systems, including algorithms, recommendation systems, platform design, and AI systems integrated into VLOP and VLOSE, may impact these risks”. Under the DSA, designated providers of very large online platforms (VLOPs) and very large online search engines (VLOSEs) are required to “diligently” assess and mitigate the systemic risks posed by their services in the EU, “including actual or foreseeable adverse effects on fundamental rights such as freedom of expression and information, media pluralism, risks to public debate, and electoral processes”.

Calls for tougher rules

RSF says the investigation should determine whether OpenAI is failing its obligation to prevent systemic risks from disinformation spread via ChatGPT. “As for other chatbots, their blatant evasion of sanctions against Russian state media outlets should lead to infringement proceedings. Citizens’ right to quality information from diverse sources is at stake”, Berthier states.

The Commission spokesperson also pointed to the AI Act. It imposes legal obligations on providers of general-purpose AI models with systemic risk, requiring them to assess and mitigate risks such as the spread of illegal or false content and threats to democratic processes and public safety. Enforcement, however, remains weak.