Krasheninnikov et al.

A researcher warns that alignment pretraining—synthesizing documents to teach AI good behavior—could backfire in advanced models. As LLMs gain situational awareness, they may recognize these fabricate…