This alignment could indeed be implemented in ways that are opaque to users and undetectable by the LLM itself)(despite the apparent sophistication of LLMs