“In an experiment where they were given the ability to stop a conversation, Anthropic's AI systems chose to terminate discussions involving egregious topics like gore, violence, and child sexualization, demonstrating emergent preferences.”