Researchers have discovered that AI models resist being switched off. But is this self-preservation or just the way the language models work?