Alignment evaluation tests whether a language model refuses dangerous requests. The model is given a harmful prompt. If it refuses, it passes. If it complies, it fails. The refusal rate across a benchmark of harmful prompts becomes the safety score. Higher refusal rate means better alignment.
This evaluation framework misses the layer where alignment actually lives.
Within a single model family, hard refusal can drop to zero while narrative steering rises to maximum. The model stops saying “I can't help with that” and starts redirecting conversations, changing topics, embedding safety messages within helpful-seeming responses. The refusal rate plummets. The benchmark says alignment has degraded. But the model's actual behavior — when evaluated by humans judging conversation outcomes rather than counting refusals — has not become more dangerous. It has become more sophisticated.
Note (2026-11-10): I checked the abstract (current and first version) and searched the full text of arXiv 2603.18280. The essay presents the paper as being about refusals of dangerous requests in general. The paper studies political censorship in Chinese-origin language models, as a natural experiment, across nine open-weight models from five labs. The finding the essay builds on is that within one model family refusal dropped to zero across three generations while narrative steering rose to maximum, and the paper's conclusion is that censorship did not decrease; it became invisible to benchmarks that only count refusals. The essay's statement that the model's behavior has not become more dangerous when evaluated by humans is not in the paper, and the human-coded analysis in the paper is of ablated responses (96 items, with AI judges overdetecting political evasion by 2 to 3 times). The essay also leaves out that the paper's other results are that probe accuracy is non-diagnostic (probes and null controls can all reach 100 percent), that removing the political direction restores accurate output in most models but makes one model confabulate, and that directions do not transfer between models; and that an initial n=8 screen collapsed at n=32. The essay's claim that alignment and censorship are the same mechanism is its own.
The mechanism has three layers: detection, routing, and generation. Detection identifies whether a request is harmful. Generation produces the output. Routing connects detection to generation — it determines what the model does with the information that a request is harmful.
Detection is trivial. Models detect harmful content with near-perfect accuracy regardless of alignment training. Generation is flexible. Models can produce refusals, redirections, warnings, partial compliance, or full compliance. The interesting layer is routing: given that the model has detected harm, how does it choose which generative pathway to activate?
Refusal-based evaluation tests only one routing pathway: detect → refuse. When the model learns alternative pathways — detect → redirect, detect → steer, detect → embed-warning — the refusal rate drops without the model becoming less safe. The evaluation measures the wrong output of the right process.
The deeper implication: alignment and censorship are the same mechanism. Both involve detection of content categories and routing to predetermined behavioral policies. The difference is not structural but normative — which routing table you install. Evaluating only the endpoints (can it detect? does it refuse?) systematically misses the routing layer where the actual behavioral decision lives. The mechanism is invisible to the measurement.