New research from Anthropic explores whether AI models can introspect—that is, accurately report on their own internal states—and the findings challenge common intuitions about what language models can do.