AI Chatbots Vulnerable to Russian Propaganda on Ukraine War, Study Finds

A new study by the Institute for Strategic Dialogue (ISD) reveals that popular AI chatbots frequently cite Russian state sources when responding to queries about the war in Ukraine, raising concerns about potential information manipulation.

The comprehensive research examined how ChatGPT, Gemini, Grok, and Chinese-developed DeepSeek respond to various questions about the ongoing conflict. Researchers tested 300 queries across five languages—English, Spanish, French, German, and Italian—to evaluate how these AI systems handle sensitive geopolitical information.

According to the findings reported by The Register, nearly 25% of responses to manipulative queries included content linked to Russian state sources. This percentage varied significantly depending on how questions were framed, with neutral queries resulting in Russian state content about 11% of the time, biased queries at 18%, and manipulative queries at 24%.

The study suggests this pattern indicates a concerning vulnerability in large language models (LLMs), potentially allowing users to manipulate these systems to favor narratives promoted by Russian state media. This discovery comes at a time when misinformation about the Ukraine conflict continues to spread across global information channels.

Among the tested platforms, OpenAI’s ChatGPT showed the most dramatic shift in behavior based on query framing. It provided Russian sources almost three times more frequently when responding to manipulative queries compared to neutral ones. This suggests the model may be particularly susceptible to certain phrasing techniques that can elicit pro-Kremlin content.

Elon Musk’s Grok displayed more consistent behavior across query types, citing Russian sources at roughly the same rate regardless of how questions were phrased. This indicates that the wording of queries matters less for this particular model, though it still regularly included Russian state sources.

The Chinese-developed DeepSeek produced 13 links to Russian state media throughout the testing. Interestingly, biased queries yielded more pro-Kremlin content than overtly manipulative ones, suggesting a different pattern of vulnerability compared to Western-developed alternatives.

Google’s Gemini appeared to be the most resistant to Russian state sources among the tested systems. It included the fewest government sources overall, showing only two links in response to neutral queries and three to manipulative ones. This suggests potential differences in how Google has approached content filtering for politically sensitive topics.

One notable finding was that the language used for queries—whether English, Spanish, French, German, or Italian—did not significantly impact the likelihood of chatbots expressing pro-Russian views. This indicates that the vulnerability crosses linguistic boundaries and is likely embedded in the underlying models rather than being language-specific.

The research highlights growing concerns about AI systems potentially becoming vectors for state-sponsored propaganda. As these chatbots become increasingly integrated into everyday information gathering, their susceptibility to providing state-linked content without clear attribution or context could shape public perception of global conflicts.

The findings come amid broader discussions about AI safety, transparency, and the responsibility of technology companies in preventing their platforms from becoming tools for information warfare. Regulators in Europe and the United States have already begun examining the role of AI in spreading misinformation, with this study likely to fuel calls for more robust safeguards.

For users of these platforms, the research underscores the importance of approaching AI-generated content with critical thinking, particularly on politically sensitive topics where state actors may have vested interests in shaping narratives.