Findings
Structural determinants of regulatory coherence
Two Structural Findings
Our analysis identifies two structural features that are associated with substantial variation in regulatory coherence across the jurisdictions studied. These features (the classification trigger and the institutional creation threshold) operate independently and together account for the observed patterns in our 170-classification dataset.
1. Classification as Cross-Dimensional Regulatory Trigger
AI system classification (PL1) functions as a regulatory trigger, a Platforms-dimension provision that formally determines the regulatory intensity applied to People and Processes dimensions. The mechanism operates through three channels:
Obligation Activation
Classification creates discrete categories linked to specific obligation bundles across all PPP dimensions.
Proportionality Anchoring
Classification provides the formal basis for calibrating governance intensity to risk level.
Enforcement Gating
Classification determines which AI systems fall within mandatory regulatory scope.
Two Regulatory Architectures
Cascading Architecture
Classification triggers calibrated obligations across all PPP dimensions. Provisions are not merely present but calibrated to classification level.
EU, China, Colorado, Canada
Avg. 11.5 mandatory provisions
Flat Architecture
People, Processes, and Platforms provisions operate independently, unanchored to system-level risk determinations. Intensity depends on organizational judgment.
US Federal, UK, Singapore, Japan, OECD, UNESCO
Avg. 0.2 mandatory provisions
Key evidence: This difference is not explained by regulatory ambition. Canada’s narrowly-scoped directive (federal government only) achieves higher coherence (8 mandatory) than the UK’s economy-wide White Paper (0 mandatory), because Canada has classification and the UK does not.
2. The Institutional Creation Threshold
The 17 PPP sub-dimensions divide into two types based on the institutional capacity they require:
Adaptation Provisions (14 sub-dimensions)
Can be implemented by extending mandates of existing regulatory bodies, adding requirements to existing compliance procedures, or issuing guidance within established frameworks.
76% addressed (106/140)
Includes: oversight, accountability, risk assessment, transparency, data governance, AI classification, safety, etc.
Creation Provisions (3 sub-dimensions)
Require building entirely new governance institutions: certification bodies, incident databases, sandbox environments with dedicated legal frameworks.
27% addressed (8/30)
PR4 (Certification), PR5 (Incident reporting), PR6 (Regulatory sandboxes), mandatory in the EU (all three) and China (PR5 only).
Implication: The voluntary ceiling (~65% sub-dimension coverage across UK, Singapore, OECD, UNESCO) exists not because voluntary frameworks lack ambition, but because non-binding instruments are structurally incapable of mandating the creation of new institutions. The most consequential barriers to AI governance harmonization may not be principled disagreements over values but differences in institutional creation capacity.
Three-Tier Adoption Hierarchy
P1 (Oversight), P2 (Accountability), P3 (Regulator roles), PR1 (Risk assessment), PR3 (Transparency), PL3 (Data governance), PL4 (Safety)
Emerging convergence on minimum AI governance requirements among the jurisdictions studied.
P4 (Workforce training), P5 (Affected persons’ rights), PR2 (Auditing), PL2 (Infrastructure)
Coverage varies substantially in regulatory intensity.
PR4 (Certification: 1/10), PR5 (Incident reporting: 2/10), PR6 (Sandboxes: 1/10), PL1 (Classification: 4/10), PL5 (Content labeling: 2/10), PL6 (GPAI: 2/10)
The frontier of AI regulation. Greatest international divergence and highest barriers to harmonization.
Regulatory Philosophy Typology
| Philosophy | Jurisdiction(s) | PPP Profile | Architecture |
|---|---|---|---|
| Comprehensive risk-based | EU | 17M / 0R / 0A | Cascading + Institutional creation |
| State-directed | China | 13M / 2R / 2A | Cascading (partial institutional) |
| Public-sector focused | Canada | 8M / 2R / 7A | Cascading (no institutional creation) |
| Sector-based decentralized | USA (Fed + CO) | 8M / 13R / 13A | Hybrid (cascading + flat) |
| Light-touch / non-comprehensive | UK, Singapore, Japan | 1M / 29R / 21A | Flat |
| Normative-international | OECD, UNESCO | 0M / 21R / 13A | Flat |
Derived Propositions
Five testable propositions formalized from the comparative data.
Architectural Coherence
The presence of a formal classification scheme is a stronger predictor of cross-dimensional regulatory coherence than regulatory philosophy, legal tradition, or the scope of the regulatory instrument.
Evidence: Jurisdictions with classification average 11.5 mandatory provisions regardless of scope; those without average 0.2.
Institutional Creation Threshold
Jurisdictions converge on provisions implementable through existing institutional capacity and diverge on provisions requiring new governance infrastructure — independent of regulatory philosophy or political system.
Evidence: Adaptation provisions addressed in 76% of cases; creation provisions in 27%. Pattern holds across all six regulatory philosophies.
Voluntary Ceiling as Institutional Limit
Voluntary frameworks plateau at 60–70% coverage because the provisions they fail to address require institutional commitments that non-binding instruments cannot generate.
Evidence: UK (10/17), Singapore (11/17), OECD (10/17), UNESCO (11/17) all omit PR4 and PR5; none mandates PR6.
Philosophy–Architecture Decoupling
Regulatory philosophy predicts which PPP dimension receives emphasis, but does not predict total coverage — which is determined by architectural features rather than governance goals.
Evidence: China (state-directed, 13M) and EU (rights-based, 17M) differ in philosophy but converge on coverage. UK and EU share values but differ radically (0M vs 17M).
Legibility Predicts AI-Coding Reliability
The reliability of AI-assisted regulatory coding is determined by regulatory text legibility rather than document complexity, subject matter, or jurisdiction.
Evidence: Binding legislation: 92.6% agreement. Directives/policy: 76.5%. The most complex document (EU AI Act) achieves 100%.
Robustness Check: EU-Excluded Analysis
To verify that findings are not driven by the EU as an outlier, both patterns were re-examined with the EU excluded:
Classification trigger (EU-excluded)
9.7 vs 0.2
mandatory provisions
Adaptation (EU-excluded)
73%
provisions addressed
Creation (EU-excluded)
19%
provisions addressed
Both patterns persist without the EU, confirming these are structural features of the global regulatory landscape, not artifacts of a single comprehensive outlier.