Measuring Regulatory Semantic Divergence and Obligation Strength Distribution Across US State Artificial Intelligence Statutes

Abstract

Lacking uniform federal artificial intelligence (AI) legislation, U.S. states have become the primary drivers of AI governance and policy, creating a landscape often described as a regulatory “patchwork.” With this characterization remaining mostly qualitative and unmeasured, this study introduces a computational pipeline to directly quantify this patchwork, using transformer-based semantic embeddings and deontic modal classification on a corpus of enacted AI statutes from California, Colorado, Connecticut, New York, and Texas, alongside comparison points from statutes before substantial amendment. Comparing the semantic alignment of statutes with the National Institute of Standards and Technology (NIST) AI Risk Management Framework reveals divergent regulatory philosophies and approaches. While California and New York focus on technical verification, Colorado’s 2024 statute featured extensive, multi-dimensional NIST alignment. Colorado's 2026 repeal and new statute reversed this stance entirely, marking a distinct pivot away from NIST ideals over time and providing quantitative evidence of the state-level retreat from federal alignment benchmarks. A sentence-level t-SNE projection of active legislation confirms this AI policy fragmentation beyond NIST alignment and focus, indicating that the substantive content of the laws themselves shows minimal semantic overlap with localized, jurisdiction-specific text clusters. Beyond the legal regions the laws covered, a Strictness Ratio reveals large variance in mandatory-language density, ranging from Texas’s discretion-heavy 50% to New York’s prescriptive 92%. Together, these findings establish a quantitative baseline demonstrating that state AI statutes diverge significantly in both linguistic structure and operational focus. By quantifying divergence of AI policy across the United States, businesses, policymakers, and citizens can further understand the fragmented regulatory landscape.