Methodology

Qualitative comparative documentary analysis with AI-assisted coding validation

Research Design

The survey employs a qualitative comparative documentary analysis structured in three phases. Twelve regulatory documents across 10 jurisdictions are systematically coded against the PPP framework, producing 170 classifications that are independently validated through an AI-assisted coding pipeline.

Phase 1

Document Collection

Assemble the primary corpus of 12 regulatory instruments across 10 jurisdictions, selected for regulatory significance, typological diversity, and geographic coverage.

Phase 2

Structured Coding

Apply the PPP coding scheme (17 sub-dimensions) to each document. Rate each as Mandatory, Recommended, or Absent with provision-level evidence.

Phase 3

Comparative Analysis

Construct a jurisdiction × sub-dimension matrix. Analyze within-jurisdiction profiles, cross-jurisdictional patterns, and interdependencies.

Document Selection Criteria

CriterionInclusionExclusion
Document typeBinding legislation, executive directives, official policy frameworks, intergovernmental standardsDraft proposals, academic commentary, media reports
Temporal scopeInstruments adopted or in force as of early 2026Expired or fully superseded instruments
LanguageEnglish originals or official translationsDocuments without authoritative English versions

Coding Procedure

Each regulatory document was coded against all 17 PPP sub-dimensions. For each sub-dimension, the coding captured:

Presence / Absence

Whether the regulatory instrument addresses this sub-dimension.

Regulatory Intensity

Mandatory (binding obligation), Recommended (encouraged but not binding), or Absent (not addressed).

Provision Detail

Specific articles, sections, or clauses corresponding to the sub-dimension.

Textual Summary

Brief description of the provision’s content and scope.

AI-Assisted Coding: Validation Experiment

All 170 classifications (17 sub-dimensions × 10 jurisdictions) were independently coded by a large language model and compared against manual expert coding. This serves as both a validation instrument and a methodological contribution demonstrating the viability of AI-augmented regulatory analysis.

87.6%

Overall agreement across 170 classifications

Agreement by Jurisdiction

JurisdictionAgreeDisagreeRate
EU AI Act17/170100.0%
Japan AI Promotion Act16/17194.1%
OECD AI Principles16/17194.1%
UNESCO Recommendation16/17194.1%
Colorado AI Act15/17288.2%
China (combined)15/17288.2%
Singapore Framework15/17288.2%
NIST AI RMF 1.013/17476.5%
Canada DADM13/17476.5%
UK White Paper13/17476.5%
Overall149/1702187.6%

Agreement by Legal Form

International standards (n=34)94.1%
Binding legislation (n=68)92.6%
Voluntary frameworks (n=17)88.2%
Directives / policy (n=51)76.5%

Agreement by PPP Dimension

People (P1–P5) (n=50)90%
Platforms (PL1–PL6) (n=60)88.3%
Processes (PR1–PR6) (n=60)85%

Disagreement Patterns (21 disagreements)

9/21

AI overrates coverage

The LLM classified provisions as present where the manual coder judged them too tangential or indirect to constitute substantive coverage.

7/21

AI underrates coverage

The LLM coded provisions as Absent where the manual coder found them indirectly present, particularly for infrastructure standards (PL2).

5/21

Intensity confusion

The LLM disagreed on the Mandatory/Recommended boundary for provisions using strong normative language without explicit penalties.

Key finding: Agreement is predicted by regulatory text legibility (the degree to which obligations specify identifiable duty-holders, enumerated actions, and explicit conditions) rather than by document complexity, subject matter, or jurisdiction. The most complex document (EU AI Act, 113 articles) achieves 100% agreement; shorter voluntary frameworks achieve 76.5%.

Methodological Limitations

  • Single-coder design constrains inter-coder reliability; mitigated by documented coding decisions with provision-level evidence and AI-assisted validation (87.6% agreement).
  • The three-level coding scale (Mandatory / Recommended / Absent) collapses substantial within-category variation. The EU’s five-article GPAI regime and China’s single-regulation approach both receive a Mandatory rating.
  • The survey captures regulatory design as of early 2026, not implementation or enforcement in practice.
  • Reliance on English-language documents may not fully capture the intent of instruments originally drafted in other languages, particularly Chinese regulatory texts.
  • The ten-jurisdiction scope excludes significant regulatory activity in Latin America, South Asia, and Africa.