Can LLMs Introspect? A Reality Check (Shauli Ravfogel)

Tea Talk June 12, 2026 Several recent works have claimed evidence of introspection in large language models, from "biofeedback" experiments where models predict latent variables defined over their own hidden states, to "steering awareness," where models report whether their activations have been perturbed. In this talk, I take a critical look at these claims through the lens of human metacognition research, arguing that establishing introspection requires both privileged access (access to information not recoverable from the input alone) and evidence of a second-order mechanism distinct from o