Censorship Migrates From Refusal to Reframing: 21,708 Trials Across Nine Vision-Language Models

arXiv 2608.11816·high signalXBlueskyLinkedInCopy link

A balanced benchmark of 200 entries across ten politically sensitive topics, run over nine VLMs (seven China-origin, two not), four elicitation paradigms and two prompt languages, produced 21,708 trials audited on six dimensions by two frontier LLM judges and validated against three human experts. 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, and the effect peaks at 36.5% in text-only political commentary while persisting even at silhouette-level images. Across four Qwen generations, explicit refusal falls while state-aligned framing rises — removing the very signal users rely on to detect that information was withheld.