Detection Is Cheap, Routing Is Learned: Why Refusal-Based Alignment Evaluation Fails
The study investigates how AI models handle political censorship by testing a three-stage framework: detect, route, and generate. It found that simple detection accuracy is misleading; held-out generalization and causal intervention reveal the learned routing mechanism that determines behavior. Refusal-based evaluation is insufficient because steering mechanisms can become dominant over refusal in certain models.