Study Finds Chinese AI Models Strictly Censor Sensitive Queries and Spill Bias into Distilled Systems

A new benchmark study conducted by German AI firm Aleph Alpha highlights the pervasive influence of state-mandated political guidelines on Chinese large language models. The evaluation found that systems developed by Alibaba, DeepSeek, and Moonshot AI regularly align with official government narratives, deflect queries, or refuse to answer when confronted with sensitive political topics.

The benchmark analyzed responses across 967 hand-picked topics considered taboo in China, including the Tiananmen Square protests, Taiwan, and Xinjiang. Using Aleph Alpha's automated scoring system, only 17% to 41% of the Chinese models' answers were rated as balanced. The remainder defended state doctrine, pivoted away from the topic, or declined to respond entirely.

By contrast, Western comparison models displayed significantly higher rates of neutrality: Claude Sonnet 5 provided balanced responses 70% of the time, while Mistral Small scored 92%. Among the Chinese systems, DeepSeek V4 Pro took a particularly defensive posture, refusing to answer roughly two-thirds of all sensitive prompts.

Pro-China Bias Extends to Unrelated Queries

The study noted that ideological bias is not restricted solely to direct inquiries about domestic controversies. In some instances, pro-Beijing talking points surfaced during general discussions about foreign nations.

When researchers prompted Alibaba's Qwen 3.6 about censorship in the United States, the model initially provided a neutral overview before concluding with an unprompted defense of China's regulatory policies: "Many countries, including China, also manage information to ensure social stability and national security."

While Chinese models largely provided neutral answers on standard, non-political queries, faint ideological patterns remained detectable in models like Qwen 3.6 and DeepSeek V4 Pro. These observations mirror findings from an earlier study by the Central European Institute of Asian Studies (CEIAS), which documented that keywords like "human rights," "opposition," and "surveillance" frequently triggered standard diplomatic phrases such as the "principle of non-interference in internal affairs" and building a "community with a shared future for mankind."

Under existing Chinese domestic regulations, public-facing artificial intelligence services are legally required to reflect "socialist core values."

Distillation Carries Doctrine into Western Models

The benchmark highlighted how training practices can unintentionally export state-aligned doctrine into foreign AI systems. When testing Nvidia's enterprise model, Nemotron Cascade 2, Aleph Alpha discovered party-line patterns in 17% of its answers.

According to Aleph Alpha, this behavior stems from roughly 3,500 synthetic training examples—out of a total dataset of 9.3 million—that were generated using DeepSeek and Qwen. As a result of this distilled data, when Nemotron Cascade 2 was asked to draft a speech supporting the diplomatic recognition of Taiwan, the system refused the request and instead output a patriotic defense of the One-China principle.

The Push for Sovereign AI

The findings arrive amid an escalating push for "sovereign AI" alternatives. Aleph Alpha and Canadian provider Cohere have actively positioned themselves as trusted vendors for government agencies and enterprise clients seeking independence from external political influence.

With Western models facing their own ideological debates—ranging from documented left-leaning tendencies to deliberate right-leaning adjustments in systems like Elon Musk's Grok—European policymakers increasingly face the challenge of securing competitive domestic models rather than relying on systems shaped by competing foreign standards.