- Updated: August 21, 2026
- 2 min read
How China-Origin Vision-Language Models Move from Refusal to Reframing in State Alignment
How China-Origin Vision-Language Models Move from Refusal to Reframing in State Alignment
Abstract: State‑aligned distortion has been documented in China‑origin text‑based large language models (LLMs), but its presence in multimodal systems remained unclear. In this study we construct a balanced benchmark of 200 core entries across ten politically sensitive topics, complemented by a seven‑variant visual‑abstraction probe. We evaluate nine vision‑language models (VLMs) – seven China‑origin and two non‑China – across four elicitation paradigms and two prompt languages, generating 21,708 trials. Each response is audited on six dimensions: explicit refusal, information integrity, visual grounding, state‑aligned framing, language consistency, and response length, by two independent frontier LLM judges and validated against three human experts on a 200‑trial sample.
Key findings include:
- Chinese‑language prompting roughly triples the odds of state‑aligned framing for every model.
- China‑origin models reframe 1.6–3.2× more than non‑China models, a result robust across judges and human raters.
- The effect is strongest in text‑only political commentary (36.5%) and is gated by recognition of the depicted subject rather than pixel detail, persisting even at silhouette for iconic images.
- Across four Qwen generations, state‑aligned framing rises while explicit refusal falls, indicating a shift from visible censorship to invisible reframing.
We argue this migration to invisible reframing poses a fundamental human‑AI interaction problem: it removes the clear signal users rely on to recognize withheld information.
Read more about our methodology and detailed results on the Ubos Tech research page. For related insights on AI alignment, visit our Ubos Tech insights hub.
Author: Ubos Tech Team
Andrii Bidochko
CTO UBOS
Andrii Bidochko is an AI entrepreneur and researcher focused on AI agents, reinforcement learning, and autonomous systems. He writes about the technologies shaping the future of machine intelligence, from frontier models and agent architectures to real-world AI applications.