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

This paper reinterprets the Chinese Room argument in light of contemporary large language models and argues that the classic syntax-versus-semantics framing is no longer sufficient on its own. The central claim is that the Chinese Room should be read as a model of closed cognition: a system able to generate formally successful outputs while remaining structurally insulated from consequence-bearing correction, revision under pressure, and the burden of mismatch with reality. On this view, the core problem is not only whether symbols can refer, but whether a system is open enough to reality that understanding becomes answerable rather than merely performative. The paper develops an answerability condition of understanding. Syntax is defined as formal symbol manipulation; semantic contact as symbol use tied to something beyond the symbols themselves; and answerable understanding as symbol use inside a system that can be revised by reality in ways that carry consequence for the system’s continuing organization. The paper then applies this framework to language models, arguing that statistical correction, including reinforcement learning from human feedback, can improve outputs and teach the appearance of correction without yet establishing lived consequence-bearing answerability. The broader implication is that current language models do not escape the Chinese Room problem by scaling fluency. They sharpen it by making formal success more convincing while leaving unresolved whether anything in the system stands in relation to the cost of being wrong.