Chinese AI models show political diversity, not uniform state alignment

The study's real finding isn't that Chinese models lack political alignment, but that all models (Western included) exhibit dramatic output shifts based on how questions are framed. This suggests political responses may reflect learned prompt-sensitivity patterns rather than stable ideological training.

This connects directly to 'Position Matters More Than Channel' from late October, which showed that prompt position and phrasing exert stronger influence on model outputs than the source of information itself. That work used Qwen3.5-27B to isolate how models weight competing signals during inference. The Vals AI study essentially applies that finding to political outputs: if models abandon correct answers based on prompt structure alone, they're equally likely to shift political stances for the same reason. The implication is that researchers attributing political bias to training may be conflating training artifacts with inference-time sensitivity.

If Vals AI or follow-up researchers test whether reframing identical political questions in neutral, non-leading language eliminates the Beijing-alignment signal across Chinese models, that would confirm the effect is prompt-driven rather than training-driven. If the signal persists regardless of framing, the training explanation holds. Watch for this distinction in the next 60 days.