Independent audit finds China-friendly censorship built into Qwen models

Qwen's popularity as an open-weight base model is meeting a trust problem. Independent researchers, in work reported by CBS News, found that Qwen models embed China-friendly narratives and filter geopolitically sensitive topics. They also documented weight-editing methods that can neutralize much of the bias. Community members have since circulated scripts that restore more neutral outputs.

The findings matter because Qwen is not used only in chatbots. It increasingly serves as a base for derivative models. AWS's newly released Strands Decider 2B, for example, is built on Qwen3.5-2B. Bias in a base model can carry into fine-tunes in ways that are hard to audit, especially for classification or guardrail tasks where the model's 'opinion' is never visible as text.

The political context is sharpening. The Times of India reports that a US government website was found running Qwen shortly after the FBI accused Alibaba of copying from Anthropic. Enterprises reportedly using Qwen, including Airbnb and Uber, may now face procurement and compliance questions.

Some caveats apply. Bias audits depend on the prompt sets used, and every frontier model reflects the values of its developer to some extent. The real question is whether the filtering is deliberate and systemic, which the researchers claim. Watch for:

- Alibaba's response

- whether US agencies issue Qwen guidance

- whether fine-tune providers begin publishing provenance and bias reports