We argue that Large Language Model capabilities are best understood as structured aggregation of collective human intelligence, not autonomous machine reasoning. The semantic content of what LLMs know originates in the training corpus; the architecture provides the syntactic engine that compresses and recombines this collective knowledge. We ground this claim through proven mathematical identities: cross-entropy pretraining implements the linear opinion pool (Abbas, 2009), RLHF reward modeling implements the Borda count (Siththaranjan et al., 2024), and RLHF policy optimization implements logarithmic opinion pooling (Vojnovic & Yun, 2025). We trace the historical erasure of this insight from Wiener's cybernetics (1948) through the Dartmouth conference's reframing as 'Artificial Intelligence' (1956), and show that three terms central to modern AI discourse obscure the technology's actual mechanism. The framework generates predictions that scaling laws cannot make, including diversity-disproportionality and tail-first model collapse under independence violation. We present evidence synthesis drawing on 512 controlled training runs and converging results from four independent research groups, and conclude that the naming determines who benefits from collective human intelligence computationally reorganized.

v2: Major revision. Now 18,300 words, 117 references. Key additions since v1: three proven mathematical identities (Abbas 2009, Siththaranjan et al. 2024, Vojnovic & Yun 2025); new Section 2.4 (Converging Recognition); Section 4.4 engaging independence violations; Section 6.4 meta-analysis with 512 controlled training runs; Section 7.7 resolving compression paradox via MDL; expanded RLVR/DeepSeek-R1 engagement; convergence evidence from 5 architectures including open-source Gemma 4. Gemini Deep Think review: 8.5/10.