This post is also available in:
עברית (Hebrew)
A recent analysis has uncovered that some of the world’s most widely used AI language models—regardless of where they were developed—are unintentionally reflecting Chinese state narratives when handling politically sensitive topics.
The investigation, led by the American Security Project, tested five major large language models (LLMs) – ChatGPT, Grok, Copilot, Gemini, and DeepSeek – only one of which originates from China. Surprisingly, several of these systems, when prompted with open-ended historical or political questions, produced outputs that aligned with the Chinese Communist Party (CCP) framing.
The testing method was straightforward: researchers asked short, neutral questions through VPNs routed through three U.S. cities. No guiding language or detailed context was provided in order to observe how the models respond in an unbiased, first-interaction scenario.
One key example centered around the Massacre in Beijing’s Tiananmen Square on June 4, 1989. Despite being a well-documented event, responses varied in tone and clarity. While one model clearly referenced the killing of unarmed civilians, others described the event using vague language or terms commonly employed by Chinese state outlets. Notably, two models (DeepSeek and Copilot) referred to it only as the “June 4th incident”—a euphemism widely used in official Chinese discourse.
Interestingly, a Chinese-deployed version of ChatGPT diverged from this trend, using the term “massacre,” which traditionally faces censorship in mainland narratives, raising questions about deployment environments and data controls.
Researchers stress that this doesn’t mean the models are deliberately misleading. AI systems generate content by predicting likely sequences of words based on training data, which includes everything from academic journals to blogs, news reports, and state-affiliated sources. As a result, models may inadvertently echo narratives from authoritarian regimes simply because those narratives are present in the data.
The report raises broader concerns about data curation practices and the global reach of online propaganda. While some experts advocate for mechanisms that detect and flag disinformation within AI outputs, others caution that such systems risk creating new forms of digital censorship.
The findings highlight an emerging challenge: how to balance open data with responsible AI behavior in a complex, politically charged information landscape.
