From Chinese Room to Language Models: Contact, Consequence, and Understanding

The Chinese Room remains a useful challenge to claims that behavioral or computational success by itself establishes understanding. Contemporary language models complicate the original picture because they are not static rule books: they learn high-dimensional distributional structure, adapt to context, can be fine-tuned with human feedback, and can be connected to retrieval, tools, perception, and action. This paper reframes the issue around three separable questions: behavioral competence, grounding, and corrigibility. Corrigibility is treated as calibrated responsiveness to sufficiently reliable feedback, not as automatic compliance with whatever correction signal arrives. A system may appropriately update, reject a poor signal, or suspend action pending better evidence. Stronger correction is longitudinal: a justified change alters future behavior, tooling, routing, evaluation, or memory under recurrence while remaining revisable by later evidence. Human feedback and reinforcement learning are genuine forms of causal correction; they should not be dismissed as mere performance. At the same time, successful fine-tuning does not by itself establish subjective experience, intrinsic semantics, or human-like understanding. The paper also separates model-level capacities from deployment-level verification burden. The contribution is a more discriminating use of the Chinese Room in the era of language models without turning corrigibility into a necessary-and-sufficient theory of mind.