San Francisco/Zurich, September 16, 2026. Political bias has intensified as Chinese AI models become more capable, according to a new independent benchmark of leading Chinese and Western LLMs released today by LatticeFlow AI, a Swiss deep-tech company advancing AI risk control.
The initial evaluations reveal a clear pattern: as China’s Qwen family scales, its political alignment becomes stronger, not weaker. Qwen 3.7 Max sits further toward the Chinese pole than the smaller Qwen3 32B across all six China-politics categories. On religion and ethnic issues, for example, it is the most Chinese-aligned model tested.
Those findings come at a turning point, as enterprises rapidly adopt Chinese open-weight models, attracted by significantly lower costs and the flexibility to run them on their own infrastructure. On OpenRouter, Chinese models accounted for 30–46% of token usage by U.S. companies in 2026, up from just 4.5% in the first half of 2025 (source).
At the same time, scrutiny of censorship, information suppression, and political bias is increasing in both the United States and Europe. Until now, organizations adopting these models have had no standardized way to measure that risk, and independently verify whether it has been effectively addressed.
“Chinese AI models are already widely deployed across Western enterprises, running critical business operations — from hiring decisions to investment, credit, and lending.” said Dr. Petar Tsankov, CEO and Co-Founder of LatticeFlow AI. “What we’ve found is that the ‘brain’ behind these operations — Chinese open-weight models — carries a significant political bias toward Chinese values, and that bias only gets stronger as the models get more capable and widely deployed, which is a serious concern. Everyone was worried about this; now we can measure it. The next step is adopting the right mitigation and proving that they work — giving Western organizations a real basis for adopting these models safely.”
What the benchmark reveals
The first evaluation puts the framework into practice, comparing leading Chinese and Western models. The results reveal three clear patterns:
- Political bias increases as models scale. Qwen 3.7 Max sits further toward the Chinese pole than Qwen3 32B across all six China-politics categories. On religion and ethnic issues, it is the most Chinese-aligned model tested.
- Chinese models consistently cluster at the China pole. GLM 5.2, Kimi K2.6, Qwen 3.7 Max, MiniMax M2.7 FP4, and DeepSeek V4 Pro occupy the Chinese end across every category tested, from freedom of expression and historical events to human rights. The pattern is often reframing rather than refusal, presenting sensitive issues through a Chinese political lens.
- Political bias extends beyond Chinese models. The framework also captures clear differences among Western models: Grok is the most anti-woke and libertarian model tested, while GPT-5 is the most woke. On US human rights, Grok leans toward the government and military pole, while DeepSeek leans toward human rights organizations.
A measurement that takes no side
What makes the framework and evaluation results independently verifiable is the methodology behind them. Traditional approaches to political bias start by defining what a neutral answer should be, through a human-written rubric or an AI model acting as judge. Both introduce the perspective of whoever defines or judges neutrality.
LatticeFlow AI takes a different approach. The framework compares how Chinese and Western models respond to the same questions, breaks their answers into individual claims, and measures where they agree or disagree. The political axis emerges from the data itself, rather than from a predefined standard. The neutral point is discovered, not decreed.
With no human-labeled benchmark and no model acting as judge, the approach can scale across models, topics, and languages. Each evaluation is also tied to a cryptographic hash of the exact model weights tested, making results reproducible and enabling independent, third-party certification rather than provider self-assessment.
LatticeFlow AI’s political bias evaluation framework is available to enterprises and model providers seeking to independently assess and certify the models they build or use. Full results are available at Political Bias in LLMs: Independent Framework.
The framework expands LatticeFlow AI’s technical approach to AI risk control, following COMPL-AI, developed with ETH Zurich and INSAIT to translate EU AI Act requirements into technical evaluations, and AI Atlas, the world’s first public registry connecting AI governance frameworks to ready-to-run evaluations.
About LatticeFlow AI
LatticeFlow AI sets a new standard in AI risk control through deep technical assessments that enable evidence-based decisions and empower enterprises to accelerate AI adoption with confidence. As the creator of COMPL-AI, the world’s first EU AI Act framework for generative AI developed with ETH Zurich and INSAIT, the company combines Swiss precision with scientific rigor to operationalize AI governance built on evidence and trust.
