⚡ Quick Summary
Published by Partnership on AI (PAI), this report examines whether current policy frameworks for foundation model documentation are interoperable and what interoperability challenges are foreseeable. It compares eight frameworks: the OECD AI Principles, the Seoul Frontier AI Safety Commitments, the G7 Hiroshima AI Process Code of Conduct, the Council of Europe AI Convention, the EU AI Act, US AI Executive Order 14110, the NIST AI RMF with its Generative AI Profile, and the UK AI White Paper and follow-up.
The analysis focuses on policy interoperability, described as the level of compatibility or consistency between frameworks, and treats interoperability as a spectrum running from identical content to direct inconsistencies. Its central finding is that documentation is a common feature of the frameworks, but they remain light on detail about the form and content of documentation artifacts, so there are no significant conflicts yet; divergence could arise as the EU AI Act Codes of Practice and further iterations of the Hiroshima Code of Conduct are developed.
The report closes with principles and numbered recommendations on coverage thresholds, EU and G7 alignment, standardization of documentation artifacts, research collaboration, and the role of AI Safety Institutes and the Network of AISIs.
🧩 What’s Covered
- Introduction and purpose: Sets out the paper's questions on whether foundation model policy frameworks are interoperable, defines documentation as information recorded for an external audience, and names examples such as model cards and datasheets.
- Interoperability and why it matters: Distinguishes policy, institutional and technical interoperability, presents a "Spectrum of Interoperability" from identical content to direct inconsistencies, and lists safety, accountability, efficiency and innovation benefits, including for SMEs and downstream developers.
- Documentation for foundation models: Explains why foundation models are hard to regulate (multi-sector use, black-box behaviour, very large training datasets, rapid innovation, downstream fine-tuning) and reproduces a documentation and accountability chain diagram adapted from PAI's Risk Mitigation Strategies for the Open Foundation Model Value Chain.
- Frameworks reviewed: Table 1 lists the eight frameworks against columns for high-level transparency commitments, documentation practices, documentation artifacts, further provisions in development, and whether foundation models are specifically addressed.
- Documentation requirements in each framework: Describes what the OECD AI Principles, COE AI Convention, Hiroshima Code of Conduct, Seoul Commitments, US AI Executive Order, NIST AI RMF and Generative AI Profile, EU AI Act and UK sectoral approach each require or recommend.
- Mapping and comparison: Tables 2A, 2B and 2C compare requirements by AI lifecycle stage; the most commonly referenced artifacts are technical documentation, instructions for use, information about datasets and incident reports.
- Interoperability analysis and thresholds: Finds no significant conflicts yet, sets out four risks and opportunities, and contrasts the EU AI Act and US thresholds of 10^25 and 10^26 FLOPs.
- Recommendations: Lists principles to adopt and Recommendations 1-5 with sub-recommendations on thresholds, EU Codes of Practice and the Hiroshima Code, standards, research and AI Safety Institutes; Table 4 compares the US NIST AISI, UK AISI and EU AI Office.
💡 Why it matters?
The report gives governance, compliance and standards teams a comparative map of what eight foundation model frameworks actually ask to be documented, and shows where the same subject matter sits in different artifacts, such as a public summary of training data under the EU AI Act against dataset documentation that need not be public under the Hiroshima Code of Conduct. It also sets out the coverage thresholds that determine which models attract additional requirements, helping organisations prepare for the EU Codes of Practice and for further iterations of the G7 Code.
❓ What’s Missing
The paper states that consideration of non-AI-focused legal requirements is outside its scope, so the tensions consultation participants raised between documentation duties and data minimisation or copyright law are named but not resolved. It offers no template for the form or content of any artifact and records that there is no consensus on best practice, particularly for dataset documentation. The framework sample is drawn from the EU, US, UK and multilateral bodies based largely in the Global North, which the authors acknowledge, and no publication date, version marking or URL for the report itself is printed.
👥 Best For
Best suited to policy analysts and government advisers tracking foundation model documentation requirements across jurisdictions, compliance leads mapping EU AI Act general-purpose AI obligations against multilateral commitments, standards participants working on dataset and technical documentation, and staff of AI Safety Institutes and similar bodies planning evaluation cooperation.
📄 Source Details
The full title as printed is Policy Alignment on AI Transparency: Analyzing Interoperability of Documentation Requirements across Eight Frameworks. It was written by John Howell and Stephanie Ifayemi and published by Partnership on AI. No publication year, version number, series or reference number is printed on the cover, title page or imprint, and no URL for the report itself appears in the text. The document runs to 41 pages; the input was a complete text extraction of all 41 pages.
