Wisevoter
World

Researchers Exposed Censorship Within Alibaba AI Model

The Qwen model has been found to prioritize Chinese government narratives and contain security vulnerabilities.

Updated on Oct. 2, 2026 in Artificial Intelligence

Isometric editorial illustration of misaligned ceramic blocks on a structural plinth, visualizing systemic bias in AI software.
Researchers identified significant political bias and narrative censorship in Alibaba’s widely used Qwen AI model, raising concerns for corporate and government users. AI Illustration. Upload story photo >

Live Poll

Should American companies be allowed to use AI models developed under foreign government censorship rules?

Researchers identified programmed political bias and censorship in the Alibaba-developed Qwen AI model. The model frequently refuses to address topics like the 1989 Tiananmen Square massacre and denies the existence of forced labor camps.

Why it matters

The model is widely used by Western corporations and U.S. government projects due to its free, powerful capabilities. However, its alignment with state-mandated narratives poses significant security and misinformation risks.

Hirundo researchers tested 500 prompts across 15 topics, finding bias in 89.8% of cases for the original Qwen model. Additionally, code generated by Qwen for U.S. projects showed 130% more security vulnerabilities.

The players

Alibaba

This Chinese multinational conglomerate is the developer of the Qwen open-weight artificial intelligence model.

Hirundo

This Israeli startup research firm specializes in analyzing and mitigating political biases within AI models.

Booz Allen

This American management and information technology consulting firm conducted research on security vulnerabilities in AI-generated code.

The details

Alibaba's open-weight Qwen model has surpassed 3 billion downloads and currently powers various functions at companies like Uber and Airbnb. Researchers have successfully modified model weights to create a Westernized version, reducing identified bias from 89.8% down to 2.8%.

Timeline

  1. June 1989 marked the military massacre in Tiananmen Square, a topic Qwen refuses to acknowledge.

  2. Late 2024 saw Chinese open-weight models holding under 2% of global usage share.

  3. June 2026 marked a surge to 45% global usage for Chinese open-weight AI models.

  4. August 2026 was the month the Qwen model exceeded 3 billion global downloads.

The Tech Race

This revelation marks a critical shift in the broader trajectory of AI development, following the pattern set by the rise of Chinese open-weight AI model usage to 45% of the global market by June 2026. The integration of these models into critical infrastructure creates a new technological arms race regarding transparency and security.

Users and developers relying on Qwen for software creation should be aware that the code may contain significant security vulnerabilities compared to non-proprietary alternatives. Businesses using the model for daily operations may need to vet outputs for political bias or alignment with state-mandated narratives.

The takeaway

The use of free, open-weight foreign AI models can introduce unexpected censorship and security flaws into private workflows. Organizations should independently audit model outputs to ensure they align with operational standards and security requirements.

Further reading

For more on the challenges facing current generative systems, visit the Artificial Intelligence section.

Source note: This article includes information reported by CBS News.

Live Poll

Should American companies be allowed to use AI models developed under foreign government censorship rules?