The Larynx Problem

Large language models have achieved remarkable fluency in producing human-like text, leading many to conflate linguistic competence with intelligence. This paper argues that this conflation is a category error with deep historical roots — analogous to mistaking the larynx for the brain. We term this the *Larynx Problem*: LLMs are trained on the output channel of human intelligence — language — not on intelligence itself. The brain does not think in words; it processes information as distributed activation patterns across functionally specialized regions, producing language only at the terminal encoding stage. We trace how this error arose from a 1969 substitution in AI research goals — from *how does intelligence emerge?* to *how do we program intelligent behavior?* — a substitution that the Transformer era made invisible by achieving unprecedented performance on its own terms. We then describe an alternative grounded in the FNC Framework and the Free Energy Principle, instantiated as the Bisociation Engine: a cognitive architecture in which language generation is one module among many, specialist knowledge domains compete in a global workspace, and creativity is operationalized as bisociation — the productive collision of structurally separated frames of reference. Empirical comparison shows that the Engine's domain-pair rankings are negatively correlated with those of a leading LLM (Spearman ρ = −0.25), confirming that the two systems optimize for structurally different properties. The destination the Dartmouth pioneers set in 1956 — machines that think, not machines that describe thought — remains the right one. The path needs correcting.