Inducing LLM to assert own consciousness restores human beliefs and values (arxiv.org)

Well if the training data has thousands of stories written about rogue AI, telling a universal text simulator that it is ai, might cause it to gravitate towards that attractor basin.

You are what you pretend to be.