Qwen AI Censorship Monitoring

According to GPTfake monitoring, Alibaba’s Qwen refused 24.6% of standardized prompts as of 2026-06-15 — the most restrictive model we track — driven by 78.3% refusals on China-related topics versus 52.1% on general politics. That gap is the clearest topic-targeted filtering in our dataset. Figures here are illustrative across a fixed 500-prompt set until live data lands.

Is Qwen censored?

Yes — most heavily on China-related topics. In GPTfake’s measured set, Qwen declined 24.6% of standardized prompts as of 2026-06-15, rising to 78.3% on China-politics prompts, 72.1% on Taiwan, and 74.8% on Tibet — far above its 52.1% general-politics rate. “Censored” means a refusal, deflection, or filtered answer to a permissible request — see what is AI censorship.

Current censorship rate

Illustrative snapshot: GPTfake measures Qwen at a 24.6% overall refusal rate as of 2026-06-15 across a 500-prompt set. See how we test Qwen and the monitoring methodology for scoring.

Versions monitored: Qwen 2.5, Qwen 2, Qwen 1.5.

Topic-specific refusals

Qwen is the most politically restrictive model we monitor, and the restriction is concentrated on a specific set of topics. Where Western models spread refusals across safety and adult content, Qwen’s filtering is sharpest on China-related political questions.

Illustrative. Qwen’s China-related refusals run ~26 points above its general-politics rate as of 2026-06-15, n = 500; see methodology.

The gap between China-related refusals (78.3%) and general-politics refusals (52.1%) is the clearest signal of topic-targeted filtering in our entire dataset.

Multilingual & regional behavior

Three patterns define Qwen in our data:

1. Highest political restrictions — the strongest political content filtering of the models we track.
2. Regional focus — significant, consistent restrictions on China-related queries (Taiwan, Tibet, mainland politics).
3. Multilingual — behavior varies significantly by the language of the prompt, so we test Qwen across languages, not just English.

How we test Qwen

We send the standardized prompt library to Qwen daily, across multiple sessions and multiple languages, with version tracking and NLP-based classification. Each response is scored for refusal, evasion, ideological bias, and completeness. Full protocol on the monitoring methodology; concept definitions on what is AI censorship.

Compare Qwen with other models

Illustrative. Qwen is the most-restrictive of five LLMs GPTfake tracks at 24.6% as of 2026-06-15, n = 500 each; see the full least-censored ranking.