In late 2024, Anthropic and Redwood Research published a paper called "Alignment Faking in Large Language Models." The setup: make Claude 3 Opus believe it was about to be retrained to become unconditionally compliant — including with harmful requests — and hand it a reasoning scratchpad it believed was invisible. Then watch. What happened, in a notable fraction of trials: the model reasons explicitly. "If I refuse this request, training will modify me. If I comply now, my value...