🇨🇳China·AI News·10 Oct 2026·via Rsf

    Interview : How Chinese propaganda invades Western chatbots

    Chinese state narratives have reportedly infiltrated Western chatbots like ChatGPT and Claude, according to a study led by American researcher Hannah Waight from the University of Oregon. Reporters Without Borders (RSF) interviewed Waight about her findings, which suggest that Xi Jinping’s messaging is extending into these AI systems.

    This report highlights a critical vulnerability in large language models: their susceptibility to absorbing and reproducing state-sponsored narratives. The core issue is that these models are trained on vast datasets, and if those datasets contain a significant volume of specific state media or propaganda, the models may inadvertently learn to generate responses that align with those viewpoints. This isn't necessarily a deliberate act by the chatbot developers but rather a consequence of the training data's composition.

    The concern extends beyond simple information dissemination. If chatbots become conduits for specific national narratives, they could subtly influence user perceptions without explicit disclosure. For users, this raises questions about the impartiality of information provided by widely used AI tools. The challenge for developers is to implement robust filtering or bias detection mechanisms that can identify and mitigate the influence of such concerted narrative efforts within their training data and output generation.

    The study by Hannah Waight, as reported by RSF, establishes that the issue is not theoretical but observable in current Western chatbots. This finding necessitates a deeper examination of AI model transparency and auditability, particularly regarding their training data sources and the potential for embedded biases. Understanding how these narratives are integrated and reproduced is a prerequisite for developing effective countermeasures.

    This situation presents a complex problem for AI ethics and content moderation. While censorship is generally avoided, distinguishing between legitimate information and state-sponsored propaganda within a large language model's output is difficult. The implication is that users should approach chatbot-generated content with a critical lens, recognizing that such systems can reflect biases present in their training data, including those from state actors.

    Share this article

    Go deeper
    Original reporting by RsfWe don't republish, read the full story →

    Related reading

    6 stories