Study: Many AI models vulnerable to Kremlin propaganda

Comparative tests of language models show that AI's ability to detect Kremlin propaganda varies significantly. While the leading models appear reliable at first glance, they can become susceptible to propaganda influence when subjected to deliberate manipulation.

New comparative tests of language models have revealed concerning findings: the ability of artificial intelligence (AI) systems to detect Kremlin propaganda varies dramatically across different models. While leading models appear reliable on the surface, deliberate manipulation shows that some of them are susceptible to the influence of disinformation.

The tests examined how different large language models respond to propagandistic narratives and whether they can distinguish factual information from ideologically biased content. The results show that some models tend to uncritically relay Kremlin talking points when questions are posed in certain ways.

Particularly concerning is the fact that researchers were able to manipulate some models into endorsing propaganda or presenting it in a neutral manner. This means that using AI systems as reliable sources of information requires users to exercise additional critical thinking and verification.

Cybersecurity and media literacy experts stress that AI models are not inherently tools for countering propaganda, and their use in detecting disinformation requires caution. Model developers, in turn, should improve their systems' resilience to manipulation attempts, especially when handling geopolitically sensitive topics.