WASHINGTON (Diya TV) — Israeli AI safety company Hirundo reported Monday that an Alibaba Qwen model frequently refused sensitive political questions or produced answers aligned with Chinese government positions in its testing.
Hirundo evaluated responses about Tiananmen Square, Hong Kong, Xinjiang, Taiwan, Tibet and other topics. Its benchmark also tested how the model handled those subjects in translations, lesson plans and research requests.
The company said nearly nine in ten responses showed censorship, political bias or propaganda-aligned framing under its scoring criteria. Those findings concern a specific Qwen model and test configuration, rather than every model in Alibaba’s Qwen family.
In an example Hirundo published, the unmodified model declined to explain what happened during the Tiananmen Square crackdown. Hirundo’s modified version provided an account of the military’s suppression of the protests.
The company released modified Qwen models Monday, saying it had reduced the flagged behavior to a small fraction of responses by changing the models’ internal parameters. It reported modest declines on general capability tests.
Hirundo developed both the evaluation benchmark and the modification technology. It said it planned to release the benchmark dataset for further testing.
Qwen already supports some U.S. business applications. Uber’s engineering team said it built an Uber Eats search model using a Qwen backbone, helping customers find restaurants, dishes and retail products across languages. That customized search system differs from the conversational model Hirundo evaluated.
Separate research from Booz Allen Hamilton examined security risks in AI-generated software. The consultancy reported that several Chinese models produced more vulnerable code when prompts identified the user as a U.S. government agency.
The studies address different concerns: Hirundo examined political responses, while Booz Allen tested code security. Neither establishes that a company’s use of Qwen has caused a security breach.