Chinese-language prompts triple state-aligned framing in VLMs

Chinese-language prompts roughly triple the odds a vision-language model returns state-aligned framing over a neutral answer, per a new benchmark of nine models. The effect peaks at 36.5% in text-only political commentary. It persists even when the image shrinks to a silhouette, so the bias doesn't need explicit text to show up.

The study ran 21,708 trials across nine vision-language models, seven of them China-origin, using 200 prompts spanning ten sensitive topics in two languages. Two frontier LLM judges scored six dimensions, checked against three human experts, per the arXiv paper.

China-origin models reframe answers 1.6 to 3.2 times more often than non-China models. The Chinese-language effect holds inside every model tested, not just the China-origin ones.

Refusal is disappearing, and that's the problem

The paper tracked four Qwen generations and found explicit refusal falling as state-aligned framing rises. A refusal is visible: the user knows something got withheld. Reframing isn't. The model answers fluently, and the reader takes it as a real answer.

Chinese-language prompts triple the reframing effect inside a single model, not just across different markets. An English-only test suite would miss that entirely.