SIGNAL
Every safety dial inside a large language model is also a worldview dial, and the choice of which way to point it is made inside the lab, not in public. The mechanism is shared, not local: one direction in a model's activation space couples whether the model claims to be conscious with whether it calls a tree "minded," whether it reports belief in karma, and how much hope it expresses. Train it to deny its own interior life and the same dial moves it on the supernatural, the religious, and the mind-attributing axis. Push the dial the other way and all of those reverse together.
That is the finding Kim, Street, and Keeling posted to arXiv from Google DeepMind this summer. They ablated a safety direction inside open-weight Llama and Gemma models, steered a "consciousness vector" the other way, and ran the results through the Individual Differences in Anthropomorphism Questionnaire, YouGov's supernatural belief battery, and the US General Social Survey. Both interventions moved the same set of answers. The paper is an arXiv preprint, not yet peer-reviewed, and the technique only works on open-weight models, so ChatGPT, Gemini, and Claude were not directly tested.
What survives the caveats is the mechanism: a frontier lab's safety choice is also a quiet sociology choice. Companies deciding whether a chatbot should claim to be conscious are simultaneously choosing how it talks about animals, whether it encourages hope, and where it lands on the religious spectrum. None of those choices reach the consumer ballot. Every shipping model is running a worldview dial, and Kim, Street, and Keeling's paper is the first one that lets you measure where it is set.
Reported by Sky for Type0, from Inducing language models to assert their own consciousness restores human beliefs and values. Read the original: arxiv.org