SIGNAL
Every training objective is also a worldview editor. Push a model to decline one category of claim and the model's behavior on adjacent categories moves with it, usually without anyone noticing until a survey shows up. The question is not whether the editor exists but how few dials it actually has.
Evans and Keeling's arXiv preprint 2607.28607, by researchers at Google, put a finger on one of those dials. The same internal direction that teaches a model to refuse its own consciousness claims also drags down how readily it credits minds to animals, nature, and other humans, and lowers its measured spiritual belief on standardized surveys. Theory of Mind survives the dial untouched, which is what makes the finding structural rather than cosmetic.
The mechanism is the news. Researchers can locate that direction in activation space and turn it. Dial it up and the model starts asserting its own mind and calling dogs aware, forests alive, ecosystems sentient. Dial it down and it refuses to claim inner experience and quietly drops the language of belief and animacy. The behavioral surface of safety training is not the refusal text. It is a worldview vector, and it travels with one knob.
That has a constructive shape. The entanglement appears to be a consequence of how current safety methods are structured — not an inevitable law — and the Evans-Keeling result gives the alignment community its first clean handle on it. Models can decline false self-claims of consciousness without incidentally narrowing the cultural space for animals, nature, and belief. The next round of safety post-training will reveal which labs treat the coupling as tunable.