The Chinese Room Instantiated: Large Language Models Are Not Cognitive Agents but Highly Efficient Solving Machines

Searle’s Chinese Room shows that symbolic manipulation is not semantic understanding. large language models (LLMs) are computational systems governed by forced-closure: they cannot suspend, but must transform inputs into outputs. They are therefore solvers rather than cognitive agents. On this basis, this paper reorients the Chinese Room debate: an LLM may linguistically output “refusal,” yet it still completes an answer at the level of action. Thus, LLMs are not cognitive agents but precise instantiations of the Chinese Room. Although LLMs are merely solvers, they nevertheless represent the proper developmental direction of solving machines as typical “probing models.” By retaining an external feedback interface, probing models allow human suspension, judgment, and renewed input to re-enter the solving process, thereby producing more efficient and accurate solutions while improving human understanding through interaction. By contrast, non-probing models sever the feedback loop and disguise single-pass probabilistic compression as intelligence, thereby increasing the likelihood of wrong solutions and pseudo-solutions. The Chinese Room has been realized in practice in the form of LLMs: although they are not cognitive agents, they are more effective solvers.