The Semantic Structure of AI Governance: A Computational Analysis of Cross-Jurisdictional AI Policy Documents

Abstract

This methodological study examines the semantic structure of AI governance within a purposive, repository-bounded corpus of 36 official policy documents aggregated into 32 jurisdictional cases. The corpus is treated as an audited analytical repository rather than as a systematic or population-representative sample. We combine multilingual sentence embeddings with document selection and segmentation sensitivity analyses, cluster tendency diagnostics, corrected language Freedman–Lane MRQAP, graph construction sensitivity, and blinded co-author calibration. Within the same transformer model, removing the 31-token overlap while retaining complete source token coverage produced little change in pair ordering (Spearman ρ = 0.995). However, grand mean centering reduced agreement with the raw baseline to ρ = 0.570, and agreement with a sentence-preserving character LSA representation was ρ = 0.473. The observed relative geometry is therefore highly concordant under some within-model perturbations but materially dependent on common direction removal and representation family. Under the primary one standard error rule, gap statistics selected k = 1 across all nine full-case representations. The tested procedures therefore provide no support for a stable multi-cluster governance typology. The primary country-level regional analysis excluded the European Union and yielded a positive within-minus-between similarity difference of 0.0444 under 50,000 node label permutations (one-sided and two-sided plus-one p = 0.00002). Network topology and community structure varied substantially across graph constructions, while blinded co-author similarity ratings showed a positive descriptive association with embedding similarity (ρ = 0.696). Overall, the findings support comparative conclusions about semantic relationships among the audited cases and about their sensitivity to the tested analytical specifications. The results do not establish unusual policy homogeneity, causal diffusion, implementation effects, or a unique governance network.

The authors' abstract, as published at the source. Systems, 2026 · DOI ↗