The Inherited Constraint: Biological Throughput Limits Shape the Information Structure of Human Language and, Through It, AI

Silicon AI has no thermodynamic throughput basin, yet language models converge on approximately 3.90 bits per token—within the biological basin. This study tests whether AI inherits its throughput from training data produced by biologically constrained brains. Bits-per-token (BPT) was measured across seven corpora using four models (Pythia-70m, Pythia-410m, GPT-2, GPT-2-large) on standardized hardware. Natural language produces BPT of approximately 3.90, near the low edge of the biological basin (centroid 4.16 ± 0.19 bits) and within about 0.5 bits of the ribosome (4.39 bits/codon). Destroying word order more than doubles per-token surprise, to approximately 10.6 bits. A granular shuffling cascade localizes the dominant structure to sentence-internal word order (syntax), contributing approximately 3.3 bits—exceeding discourse (1.3 bits) and sentence ordering (0.6 bits) combined. Zipf exponents are identical in original and shuffled English (α = −0.843), indicating that word-frequency statistics are insufficient to explain the approximately 3.90-bit convergence. Structured text costs 20% more energy per token than shuffled text. A four-link causal chain is proposed: physics constrains biology, biology constrains cognition, cognition constrains language, language constrains AI. Links 1–2 are established; Links 3–4 are supported by convergent evidence with four falsifiable predictions offered.