LLMs show a “highly unreliable” capacity to describe their own internal processes

In the end, though, the research finds that current AI models are “highly unreliable” at describing their own inner workings and that “failures of introspection remain the norm.” Anthropic’s new ...