The sixteen themes of the AAIF governance working group, read against binding law

About this documentUpdated 2026-09-21ShowHide

Sean McDermott, Co-Founder and CEO, UnGovr

Written by Sean McDermott (with AI assistance) using the LexLint law library, which supplied every legal instrument, status and date on these pages.

Every law named here links to its summary page on lexlint.org, translated to English (if needed) and restructured to a standard format for human and code use. Every case links to the court's or the regulator's own record where one could be reached.

© 2026 UnGovr, publishing as LexLint. The text and the figures are licensed under Creative Commons Attribution-ShareAlike 4.0: share and adapt them, including commercially, with credit to LexLint (UnGovr) and under the same licence. Please contact LexLint at hello@ungovr.org to discuss other terms. Logos and wordmarks belong to their owners.

Corpus figures as of 2026-09-21.

Legal information, not legal advice. This document describes the law as written and dated; it does not apply it to any system. The notice at the foot says what that means.

The Agentic AI Foundation's Governance, Risk and Regulatory Alignment working group extracts requirements from frameworks under sixteen themes. This page reads the same themes against the LexLint software-law corpus: how each of the corpus's twenty obligation classes maps onto them, what the distribution says about where statutes speak and where they do not, and which handbook documents carry the material behind each theme and each of the group's three deliverables.

1What this is

The Agentic AI Foundation (AAIF), a Linux Foundation foundation, runs a Governance, Risk and Regulatory Alignment working group whose charter sets three deliverables: a landscape of agentic-AI policy frameworks, a gap analysis, and distilled recommendations. Its monthly reports record the method it chose: about a dozen common themes, and extraction prompts to map frameworks, standards and legislation onto them consistently. In August 2026 the group circulated a workbook to its members with sixteen such themes, each with an extraction prompt to be run against one framework document at a time, and a reference guide of thirty frameworks to run them against. The workbook is not public; the themes are summarised here in the group's own words.

The prompts share one contract: classify every provision by type, decompose compound statements into atomic requirements, keep the exact modal language, and record for each requirement an identifier, the text, the type, the obligation strength, the applicable entity, the source clause and a confidence. Two of the prompts' rules matter most: do not infer applicability beyond what the source states, and say "legal interpretation required" where legal analysis is needed.

This page reads the same sixteen themes against a different source. Instead of thirty frameworks, the input is binding law by jurisdiction: the 6,825 requirement lines the LexLint software-law corpus carries for every AI, privacy, scraping and cybersecurity instrument it holds, each one already atomic, already tied to a citation, and already dated. The corpus has an axis of its own for those lines, twenty obligation classes recorded on each instrument by the research, and the requirement register in the handbook is keyed on it. The bridge between the two is a rule: a line joins a theme when its instrument's class is one the theme's prompt is about, or when the line's own wording matches the theme's key concepts. Section 2 prints that bridge in both directions, section 3 says what the result shows, and sections 4 and 5 say where in the handbook the material behind each theme and each deliverable sits. The sixteen themes, one by one holds each theme's counts and a sample of its lines.

2The crosswalk

One row per obligation class, largest first. The third column names the theme the class feeds outright, where a theme's prompt is about that class; the last column names the other themes that admitted at least a quarter of the class's lines by wording, which is where the overlap between the themes shows. Since the theme rules are generous, a line sits in 70% of cases in two or more themes, and a theme's count is an upper bound on what a careful reader would keep. 510 lines fall under no theme at all; they are in the handbook's register and its file, and absent from the theme file.

The same bridge from the themes' side: for each theme, the classes its lines carry, largest first, and how many of its lines come from an instrument with no class recorded, which is the share of the theme that rests on wording alone.

3What the distribution says

Four themes carry most of the binding law: data protection and privacy (2,791 lines in force), security and access controls (2,524), transparency and USER disclosure (2,224) and recordkeeping and auditability (2,241). These are the themes where a statute, rather than a framework, is the source of the requirement, and they are where a gap analysis against binding law will find the most to measure.

Three themes are thin, and the thinness is the finding. Architecture has 39 lines, testing and red teaming 20, monitoring and logging 168. Binding law rarely tells an OPERATOR how to structure a system, how to test it, or what to log in what format; it tells the OPERATOR what outcome to secure and what record to be able to produce. Those three themes are framework territory, and the frameworks in the group's reference guide are the right source for them. What a statute adds is the outcome the framework control has to serve, which is why the recordkeeping theme is large while the logging theme is small.

Use-case classification and risk tiering (924 lines) is dominated by prohibitions: the law's way of tiering is to name what may not be done at all. Human oversight (150) and action management (137) are small but spread across many jurisdictions (98 and 91 respectively), which is the pattern of a duty that arrives through privacy law's automated-decision rules rather than through an AI statute. Incident response (1,000) is almost entirely breach notification, the oldest and most uniform family in the LexLint software-law corpus.

The themes and the classes are not two names for one thing. A theme is a family of controls an organisation might adopt; a class is a family of duties a statute imposes. Where the two coincide, as with breach notification and incident response, the crosswalk is one row. Where they do not, as with architecture, the theme has almost no law under it and the class has no theme that is about it, and each side of the table says so in its own terms.

4Where the handbook reads each theme

The handbook at lexlint.org/agents is written for people building and running agents and is organised by their questions, not by the themes. This table says, for each theme, which of its documents carry the material a reader mapping that theme would want, and why. Each entry is the document's number in the handbook and its short name.

5The three deliverables, and what is published against each

The charter names three deliverables. For each, the handbook documents that bear on it, with the reason each is named.

6The files

Three files are cut from the same lines on the date in this page's byline, and all three will be regenerated when the corpus moves.

• theme-crosswalk.csv: the two tables in section 2 as one file, a row per obligation class with its counts, the theme it feeds outright, and how many of its lines each of the sixteen themes admits, with a last row for the lines whose instrument records no class.
• theme-register.csv: every requirement line that falls under at least one theme, one row per line, with the theme identifiers first and then the columns the prompts' contract asks for, plus the party, hook and tier columns the corpus adds. Its columns are described on the sixteen themes, one by one.
• requirement-register.csv: the handbook's register, every line on the corpus's own axes with the obligation classes and no theme columns, described in the requirement register.

theme-crosswalk.csv is licensed separately from the text and figures around it, under Creative Commons Attribution-NonCommercial 4.0 rather than the ShareAlike terms in the byline: use it for noncommercial purposes with credit to LexLint (UnGovr), and contact LexLint at hello@ungovr.org to discuss commercial use. The other two files state their own terms where they are offered: theme-register.csv in section 3 of the sixteen themes, and requirement-register.csv in section 6 of the requirement register.

7What this page does not claim

It does not say that any line binds any particular organisation. That is the applicability question, and it is answered per deployment from the roles a system holds and where its parties are, which is what the first documents of the handbook are about. The theme assignment is a rule applied to text, reviewed in samples, not a finding made line by line, and the crosswalk is a description of where that rule put the lines, not a proposal that the themes or the classes be changed. Every entry, on these pages and in the files, links to the source the corpus researched it from, so that the rule can be checked against the law.