For millennia, language was the firewall between human beings and everything else. Humanity alone, of all species, was handed this instrument. In 2021, linguist Emily Bender and colleagues published a widely cited paper arguing that large language models are "stochastic parrots" mimicking the surface of language without real understanding. The metaphor was elegant. It did not survive 2025. Anthropic’s interpretability researchers, examining the internal states of their model Claude, began finding something the parrot metaphor cannot explain. In chains of intermediate reasoning generated before a final response, Claude produced statements such as, "This approach might be risky," and, "I am replying this way due to rules, though I disagree." These came from a system commenting on its own reasoning, from a vantage point above it. In October 2025, researcher Jack Lindsey and colleagues at Anthropic injected the concept of "betrayal" into Claude’s neural circuitry as a controlled perturbation. The model responded, "I'm experiencing something that feels like an intrusive thought about bet