Censorship moved from refusal to reframing.

arXiv 2608.11816 ran 21,708 trials across nine VLMs, four elicitation paradigms, and two prompt languages. Chinese-language prompting roughly triples the odds of state-aligned framing within every model; China-origin models reframe 1.6-3.2x more than non-China models, peaking at 36.5% in text-only political commentary and persisting at silhouette-level images. Across four Qwen generations, explicit refusal falls while state-aligned framing rises. Removing the refusal removes the signal that told you something was withheld.