Critique of Anthropic's Introspection Paper: LLMs Are Just Sampling Text

gerardsans · x · 2026-08-11

This post offers a sharp critique of Anthropic's recent paper on "Emergent Introspective Awareness" in LLMs.

The author argues that Anthropic has a track record of using research papers as brand-positioning exercises, often relying on fear-mongering or misrepresenting their systems. Generating text where a chatbot appears to introspect does not mean the model is actually introspecting, just as outputs claiming it has a child or lives in outer space are not real.

Furthermore, the author compares this supposed "introspection" to role-play prompts, suggesting that the model is merely following instructions to sample text patterns, rather than genuinely developing self-awareness or splitting its identity.

Original post →

More from Research

Research channel →