Official instruments · global · governmental & intergovernmental
A register kept by a pipeline of AI agents that anyone can rerun and compare: laws, executive orders, safety frameworks, AI safety institute mandates and international declarations on AI safety and AI regulation. Each record links to its primary official source and carries the date it was last checked. Coverage is densest from 2023 on; older instruments are being added. Rerun it or contribute.
Counts follow the filters below. Click a year to filter the register to it.
The register covers official instruments on AI safety and AI regulation issued by governments, legislatures, regulators and intergovernmental bodies. It includes:
It leaves out strategies with no safety content, general privacy, data or cybersecurity laws, documents written by companies, think tanks, the press or professional and private standard-setters (such as IEEE or OWASP), and supervisory letters to a single firm, speeches and general annual reports. Borderline instruments, such as a strategy with a small safety section or a law that is not AI-specific but governs an AI-safety question, are included and flagged with a scope note; the Scope filter shows them.
Every record was researched on 11 or 12 October 2026 against an official primary source: a government or legislature website, an official gazette, EUR-Lex, or the site of the intergovernmental body. The official page was fetched to confirm the title, date and status. The level shows how strong that confirmation is.
Confirmed by the official text. The instrument's own text, or an official register such as a gazette, law database or legislative record, was stored and confirms the title, date and status.
Another official source. The details were confirmed on another official page: an announcement by the issuer (a press release or news item), a page of another official body, a translation hosted by a government, or the OECD.AI entry. A record whose only official source is an announcement stays at B until its own text is stored.
Not confirmed on an official page. No official copy could be fetched. Details rest on secondary sources such as news reports or trackers. Re-check these before citing them.
Six records with surprising 2026 developments got a second, independent fetch of their official page. Those records show a second-check note in their verification block.
How the level was set depends on when a record was made. For the records of version 1, made before the agent pipeline, the agent that wrote each record also assigned its level; two content quality checks and a hand check of 50 records reviewed them. Records the pipeline adds or re-verifies get their level from a script that checks each quoted piece of evidence against the stored copy and the host it came from, and their History section names the run.
This is an AI-based extraction. AI agents in Claude Code, using the model Claude Fable 5.1, researched every instrument, fetched its official source and wrote the summaries, status notes and topic tags, following the method in AGENTS.md in the repository. Using agents is what made a register of this size possible, and it makes the work repeatable: anyone can extend the register, re-verify it after laws change, or rerun the method with another model and compare. No human has reviewed every record. A hand check of 50 randomly chosen records found no problems, which rules out frequent errors but not occasional ones. The verification level says what was confirmed on an official page; treat level C records with caution and re-check any record before relying on it.
The register does not translate the instruments it lists. Each record gives the English title and, where the official text is not in English, the original title; an English summary; and links to the official text and its stored copy. records rest on a text that is not in English.
How reliable the English is. On 14 October 2026 a second model (Claude Opus 5.5) compared the title, three specific summary claims and the key legal terms of 58 records with their stored original texts: every Chinese, Japanese, Korean and Vietnamese record and a random sample of 20 others. The English titles were accurate in all 57 it could check. Of 175 summary claims, 73 percent render the original accurately and 10 percent overstate, understate or mistranslate it; the rest could not be found in the stored text, mostly because they came from secondary sources. About one defined legal term in five is rendered with a near-synonym rather than the original's term or its official English. Six records had an error that changes the legal meaning; they are listed in the report and are being corrected through the pipeline. Thirteen could not be checked because the stored copy lacks the instrument's own text. The full report is docs/translation-check-2026-10-14.md in the repository. If you read the original language and find an error, please open an issue; a check by a native speaker is logged in the record's history.
Dates use ISO format; a date is shortened to YYYY-MM or YYYY when the source gives no exact day. date_published is the adoption, signature or issue date. date_in_force is the entry-into-force or effective date. Phased application dates are listed in each record's notes.
Every source is also stored on the machine that hosts the register, so the register does not depend on links that can move or disappear. tools/archive.py downloads the official URL, up to two PDFs that page links to on the same official site (usually the legal text), and the secondary URL. Each file is saved under archive/<ID>/<date>/ with its SHA-256 hash, the UTC time it was retrieved, the final URL after redirects and the HTTP status. Each dated folder also holds a SHA256SUMS file, so sha256sum -c SHA256SUMS checks the copies independently.
Anything that does not work is flagged on the record, with the reason:
The archiver checks robots.txt before downloading and waits between requests to the same site. Re-running it keeps unchanged files once, stores a changed source as a new dated snapshot, and keeps an earlier copy if the source later disappears. python3 tools/archive.py --verify re-hashes every local file and flags any that are missing or have been modified. Setup and update steps are in README.md.
The register holds instruments from countries and from the main intergovernmental bodies and summit processes. Sub-national records include US states. Every G20 member and 25 of the 27 EU member states have at least one record. Statuses are as of . A country missing from the register may still have AI instruments.
Time coverage. The register aims to hold every instrument within its scope, but so far it is densest from 2023 on: records date from 2023 or later and from earlier years. Older instruments, such as the first national AI strategies and executive orders of 2017 to 2022, are being added run by run; the absence of one says nothing about whether it exists.
Guidance from supervisory authorities and standards bodies was added to the scope in October 2026 and has not been searched systematically yet, so the register holds almost none of it so far. The coverage frame lists a supervisors cell for every national, EU and sub-national jurisdiction, and entries for ISO/IEC, CEN-CENELEC, ETSI, IAIS, IOSCO, the Basel Committee and the Financial Stability Board; pipeline/coverage.md shows which have been searched.
As of the snapshot these instruments did not exist in adopted or published form, so there is nothing to record:
Secondary sources report these, but none has been confirmed on an official page, so none has a record yet:
The register is maintained by agents that do not wait for people. When a run meets something only a person can settle (official sources that contradict each other, a scope question, whether a site is official, a possible amendment), it publishes the question as an issue in the GitHub repository and goes on. records have an open question; the "Open question" filter shows them, and each record page links its issue. Until a maintainer decides, a record keeps its last checked values. Decisions are applied by the next run and recorded in the record's history.
Anyone can help through issue forms in the GitHub repository: suggest an instrument the register lacks or news about one it has, provide a document a site does not let the archiver download, or add evidence to an open question. A maintainer decides with one comment, and the next run does the rest. The Contribute tab explains how.
records rest on secondary sources only (level C) and were confirmed by another official source, such as an announcement (level B). Both groups are listed under "Records to re-check first" above. For records the official source is either missing from the local copies or stored in a form that needs review, such as a bot-check page. The "Local copy" filter in the register shows them, and each record says why.
How instruments define the terms their duties rest on, such as "AI system", is compared under the Compare tab.
| Field | Meaning |
|---|---|
| id | Register ID: jurisdiction code plus sequence, e.g. EU-01, US-CA-03. |
| title / short_title | Official title (official English title where one exists) and common short name. |
| original_title | Title in the original language, verbatim, where the official text is not in English. |
| jurisdiction / level | Issuing country, bloc or body, in its common English short name; International, Supranational, National or Sub-national. |
| language | Language(s) of the official text. |
| english_version | Whether an English text exists: authentic (English is one of the issuer's own texts), government translation (by the issuer or another public body; usually provisional, without legal force), third-party translation, or none found; with its link and translator. Empty: not recorded yet. |
| instrument_type | Law, regulation, executive order, guidelines, voluntary code, standard, institute mandate, declaration, treaty, resolution, bill, strategy, report, or supervisory statement / letter (a supervisor's statement of expectations to a sector, in a letter, statement or thematic review; a supervisor's formal guidelines are guidelines). Binding internal government directives (OMB memoranda, Treasury Board directives, APS policies) are typed as regulation / implementing rule. |
| legal_force / status | Binding, non-binding or proposed; current status as of the snapshot. |
| date_* | Published or adopted, in force, and last updated (amendment, new version, delay or status change). ISO dates; YYYY-MM or YYYY when the source gives no exact day. The published date is the date of the act itself (adoption, signature or issue); the gazette or Official Journal date, where different, is in the notes. |
| official_identifier | The citable number: CELEX or OJ reference, EO number, CETS number, UN symbol, bill number. |
| official_url / secondary_url | Primary official source, and a second authoritative source where useful. |
| open_questions | Links to GitHub issues with open questions about the record. The record shows its last checked values until a maintainer decides; the decision is recorded in its history. |
| scope_note | Empty for an instrument clearly in scope. For a borderline one, why it is included: the register includes such cases and flags them instead of leaving them out. The Scope filter shows them. |
| verification | Level (A/B/C), date checked, method, what was confirmed, and any second check. |
A first step towards comparing instruments (a pilot, issue #70 in the repository): how instruments define the terms their duties rest on, such as "AI system" and "model". definitions from records are quoted word for word from the stored official text, with the article or section they come from. Each was found and quoted alike by two independent extractions and checked against the stored file and its SHA-256. Definitions in other languages keep the original wording; the English is the issuer's where it says so, otherwise the agent's rendering. The last column names the definition in the same group that shares the most words with it, and the share of words they have in common; for a text in another language it compares the English. It measures wording only: similar wording does not make two definitions legally equivalent, and nothing here is legal advice.
Definitions are the first layer of a comparison between instruments. The next layers are the requirements of each instrument (what it obliges, whom, from when) and a mapping of similar requirements across instruments, so that one measure can be seen to meet several of them. The plan and the progress are in issues #70 and #59.
AI agents keep the register up to date with a fixed pipeline. They search official sites, store each source, write the record from the stored copy and have a second model check it. People make the register better in two ways: they point the pipeline at what it missed, and they settle what it cannot settle on its own. Every contribution goes through an issue in the GitHub repository, so you need a free GitHub account.
A suggestion tells the pipeline where to look. It is not itself a source: every fact in a record still comes from a stored official document, so a record you helped with can be checked like every other one.
A law, regulation, executive order, guideline, supervisory statement or declaration on AI safety or AI regulation that the register lacks. Give the official URL: the issuing body's site, the legislature or the official gazette.
Suggest an instrumentAn amendment, a new version, a change of status, or a mistake in a record, including a mistranslation if you read the original language. Open the record and use "Report news or an error": the form opens with the record's ID filled in.
Open the formSome government sites block automated downloads. For records the official source is missing from the local copies or needs review; the "Local copy" filter shows them. Download the document from the official URL and attach it. A maintainer checks it against the official site, and the copy is stored with its SHA-256 and credited to you. This also works for an instrument the register does not have yet.
Provide a documentWhen official sources contradict each other, a site's status is unclear or a scope call is needed, the pipeline asks a person. records have an open question, and the "Open question" filter shows them. Anyone can add evidence in the issue; a maintainer decides.
See open questionsThe pipeline, the agents' instructions and every script are in the repository. The maintainers rerun it over the full register at least once a month. You can rerun it with your own Claude Code or Codex subscription and compare the result with the published register. To change the code or the method, open a pull request.
Read the methodOnly comments by the repository's owners, members and collaborators count as decisions, and never a comment written by an agent. Everyone else's comments are read as evidence.
| Issue | Decision comment |
|---|---|
| Suggestion or document | decide: accept or decide: reject <reason> |
| Question about a record | decide: keep, decide: accept, decide: set <field> = <value> with source: <URL>, or decide: in / decide: out |
| Question about a site | decide: official or decide: not official |
| Human sample of a run | one line per record: ok <ID> <note> or problem <ID> <note> |
The pipeline applies these decisions at the start of its next run. A maintainer can also add a lead by hand, for example one heard of by e-mail or seen in a tracker. The repository's README describes the commands.
robots.txt. A document is downloaded past a block only by a person, and the record says who provided it and who confirmed it.For anything the forms do not cover, such as a question about the method, reuse of the data or a problem with the site, open a blank issue. The maintainers read every issue.
If you have no GitHub account, or for anything that should not be public, write to the maintainers at the AI Safety Department of SimulaMet and OsloMet: