Methodology

How country scores are constructed, what the numbers mean, and what this tool can and cannot tell you.

What a score means, and what it does not

Every score runs from 1 to 5 and belongs to one of two lenses.

• Implementation (the implementation index, regulation status, policy lever and enforcement level): how much AI governance is in force and operating in a country. A 1 means no observable activity. A 5 means every element of a demanding anchor is met at the national level: the rules apply today, the bodies are staffed and using their powers, and enforcement is routine and published. Anything deferred or not yet used holds a score at 4, so 5s are rare.
• Governance style (governance type and actor involvement): how a country governs AI. Governance type runs from centralised (1) to distributed (5); actor involvement from narrow (1) to broad (5). Neither end is better.

A higher score is not a better score. An implementation score says how much is in force, not whether the rules are good, fair, proportionate or effective. A country at 4.5 has more in force than a country at 2; the map does not say that it regulates AI better, and it does not rank countries. A governance style score describes a structure, not an achievement.

Every score can be traced: each dimension is the mean of four sub-indicators, each sub-indicator carries a one-sentence rationale, and every entry lists its sources and a confidence rating. The sections below set out each level and how the scores are produced. In the app, the How to read this map guide (the "?" button in the header) says the same in brief.

What this is

The AI Regulation Map visualizes the current state of AI governance in 196 countries across six dimensions, each scored 1–5. It is intended as a reference view for policy researchers, academics, and civil society: a place to form a comparable mental model of how different jurisdictions approach AI, and a jumping-off point to the primary sources behind each score.

Country shapes come from Natural Earth's 1:110m world atlas, which has no shape for about thirty small states, among them Singapore, Malta, Bahrain and most Pacific and Caribbean island states. The map draws each of them as a small circle on its main island (for Kiribati, Tarawa), coloured and selectable like any other country.

The project was originally seeded from the EA Forum AI Governance Tracker. Since then every country has been independently re-researched and re-scored by the automated pipeline described below; the current dataset no longer derives from that source.

The six dimensions

Each country is scored on five dimensions, plus a composite implementation index shown by default. In the app they appear in two groups. Three dimensions (regulation status, policy lever, enforcement level) are the implementation dimensions (normative in the rubric and the code): a higher score means more is in force and operating, not that it is better. Two (governance type, actor involvement) are the governance style dimensions (descriptive): they record how a country governs (centralised vs. distributed, narrow vs. broad participation), not how well. Governance style is excluded from the composite.

Since June 2026, each dimension score is the mean of four concrete sub-indicators, each rated 1–5 against a written definition (for enforcement level, for example: sanctions framework, actions actually taken, dedicated authority, monitoring practice). This produces quarter-point decimal scores that are traceable to specific facts; the full sub-indicator table per country is published in subscores.json. Under rubric v3 (September 2026), each level of the 1–5 scale describes an observable state that a jurisdiction can be verified to be in (see Levels below); scores researched before the first v3 run use the v2 calibration (see Rubric versions). Countries scored before June 2026 carry integer scores until their next research pass.

From methodology v2.1 (late September 2026), each research pass also gives every sub-indicator a rationale: one sentence, at most 200 characters, that states the fact the score rests on and names the instrument, body, or date where one exists. Rationales are model output, not editorial text. They are produced by the same research call that assigns the score, are not reviewed or edited by hand, and can be wrong in the same ways a score can be wrong. Treat a rationale as the claim to verify against the listed sources, not as a verified source itself. Rationales appear under each sub-indicator in the country panel and in subscores.json; a country last researched before v2.1 has none until its next research pass.

Insufficient evidence is not a 1. From rubric v3.1 (September 2026), a sub-indicator can be marked insufficient evidence when no source the research found confirms either the presence or the absence of what it asks about; its rationale then says what was searched. A 1 needs positive evidence of absence: a source that says the thing does not exist, or a well-covered country whose official record was checked and where nothing exists. Sources are thinner for some countries and languages than for others, so without this value a country the research could not see into would drift to the floor of the scale. A dimension with two or more insufficient sub-indicators has no score; with one, it is the mean of the other three. The implementation index needs at least two of its three dimensions scored. Any unscored dimension sets the entry's confidence to low. The map paints such a country in a neutral grey distinct from "No data" (no entry at all), and scores.csv leaves the cell empty.

Levels

Every sub-indicator is scored on one ladder of observable states. Each row gives the short label the country panel shows for an implementation sub-indicator, then the level's definition, quoted from the model's research prompt in prompt.py.

A 5 is not a ranking and not a judgement that the rules are good: it means the anchor is met in full, by the country's own record, whatever other countries do. It is demanding on purpose, and a single instrument never earns it. Three further rules apply to the implementation dimensions (from rubric v3.2):

• National level only. For an EU member state, EU regulations and directives that apply in it count as its national law. Laws, agencies and actions of states, provinces, regions or cities never raise an implementation score; they are recorded under governance type (sub-national role) and in the text. A US state statute, however strict, does not count toward the United States' implementation scores.
• AI-specific means written for AI. General law with no provision written for AI (data protection, consumer protection, competition, product liability, export control, cybersecurity) scores at most 2, even when a regulator applies it to AI firms.
• A narrow instrument stays narrow. A binding rule for one use case or one kind of content (deepfakes, intimate images, election content, one type of chatbot) scores at most 3 on binding force and at most 2 on scope.

The research scores the state that sources dated on or before the research date confirm, not the direction of travel and not the gap to other countries, so a country's score changes only when its own record changes. When the evidence supports two adjacent levels, the lower one is given; this tie-break never turns missing evidence into a 1. Announcements, ambition and the volume of activity are not scored, only what a sub-indicator's definition asks about.

Each sub-indicator also has its own anchors, listed under the dimensions below. On the implementation dimensions the prompt defines all five levels for each sub-indicator (since rubric v3.2); on the governance style dimensions it defines a 1, a 3 and a 5, and a 2 or a 4 sits between two anchors. The model is told to score strictly against these definitions, and they are quoted verbatim, as the model reads them. The code name beside each sub-indicator is its key in subscores.json and in the API's subscores column. On the governance style dimensions a level is a position, not an amount, so the country panel shows the anchor (or the two anchors a score sits between) in place of the labels above.

Regulation Status

How much binding, AI-specific regulation is in force: whether rules bind, how far they reach across sectors, how far they are in operation, and whether they are written for AI. Not a judgement of whether the rules are good.

Illustrations, not definitions. A cross-sector AI law in force whose main obligations apply from a later date scores 5 on AI specificity, 4 on binding force and scope, and at most 4 on implementation; this is where the EU AI Act sits for its member states until its Annex III high-risk obligations apply. A country whose only national AI-specific binding rule targets one use case scores 3 on binding force and 2 on scope; a country whose only instrument is a published national AI strategy sits near 2. A country with no AI-specific policy activity scores 1 once its official record confirms the absence.

Policy Lever

How many kinds of policy instrument are in use: binding law, guidance and standards, funding and other economic tools, and the institutions that carry them. It counts instruments in use, not whether they work.

Governance Type

Whether authority over AI is held by one body or spread across regulators, levels of government, courts and independent bodies. Descriptive, not a quality scale: a 1 here means highly centralised, not bad; a 5 means highly distributed, not good. A centralised single-authority system scores low on these sub-indicators even when it is highly effective. A 1 describes a single authority, not the absence of governance: where every implementation dimension is 1, the country panel shows “No AI governance activity observed” instead of a governance type score. This dimension does not enter the implementation index.

Actor Involvement

How widely industry, civil society, academia and international partners take part in AI policy-making, from narrow (1) to broad (5). Descriptive, not a quality scale, and scored on domestic process: international activity does not substitute for civil-society access at home, so a government that excludes domestic civil society scores low on civil society however active it is abroad. This dimension does not enter the implementation index.

Enforcement Level

How much enforcement of AI rules is observed: the penalties defined, the actions taken, the authority that owns enforcement, and the monitoring it carries out. It measures enforcement activity, not whether it is fair or effective.

Implementation Index

The implementation index is the arithmetic mean of the three implementation (normative) dimensions: regulation status, policy lever, and enforcement level. It is labelled Implementation Index on the map, exported as the Average Score column, and served as avg_score by the API. It measures how much AI governance is in force and operating, not its quality. Governance type and actor involvement are excluded because they describe how a country governs, where neither end is higher on any scale of extent; averaging them in would move the index up or down for being centralised or distributed. Shown by default on the map because it gives the fastest cross-country read, but the individual dimensions are where the interesting signal lives.

Name history: until September 2026 this composite was called the maturity index. It was renamed because readers took a high "maturity" as praise; the formula and the data did not change.

Peer sets

The country panel has a Compare with row. Each chip opens a side-by-side comparison of the selected country with a set of peers. Three kinds of set exist. Each lists the closest countries first and is cut to the comparison limit of four countries, so a set holds the selected country plus three peers. Countries without the scores a set needs are left out. Ties resolve alphabetically.

• Bloc peers: for each bloc the country belongs to (EU, G7, G20, ASEAN, African Union, BRICS+, NATO, OECD), the other members closest to it in implementation index.
• Similar implementation: the countries whose implementation index is closest to the selected country's.
• Similar profile: the countries at the smallest Euclidean distance across all five dimensions. The dimensions carry equal weight, including the two governance style ones. The aim is a similar regulatory shape, not a similar level; the implementation set already covers level.

The sets rest only on bloc membership and score distance. No other signal (region, language, income group) is used.

Confidence levels

Each country record carries a confidence label that reflects how much public, primary evidence was available to the research pass.

• High: scores are supported by enacted legislation, official regulator publications, and at least two independent English-language sources published within the last 12 months.
• Medium: scores rely on strategy documents, draft bills, or reputable secondary reporting; some dimensions may be inferred from adjacent policy areas.
• Low: public information is sparse, in a non-English primary language with uncertain translation, or predates the last major policy cycle. Treat these rows as indicative rather than definitive.

The research model proposes the label; the pipeline caps it in two cases. A record with no citable source, or with a dimension left at insufficient evidence, is low. A record is high only when at least one cited source is official (a government, legislature, regulator or EU institution site); a record the model rated high on law-firm, tracker or news sources alone is published as medium. Cited links that are dead when the pipeline checks them (not found, or a known-bad address pattern) are dropped before publication and do not count.

Confidence is currently assigned at the record level (per country), not per individual dimension. We consider per-field confidence a future enhancement.

On the map, low-confidence countries are hatched with fine diagonal lines over their score colour, so an indicative reading does not look as settled as a well-sourced one. The hatch does not change the colour, and it can be turned off under Filter, "Show uncertainty". Countries with no data keep the plain no-data fill. When the timeline shows a past date, the hatch follows the confidence recorded for that date if the history has one (the history records confidence from September 2026), and otherwise the current rating; the legend says which.

Data update cadence

A GitHub Action runs every Monday and re-researches every country with web search enabled. Manual overrides (single-country re-research, forced refresh of the whole dataset) are possible via scripts/update_data.py.

Every run's score changes are preserved as snapshots in public/history.json, which drives the timeline slider on the main page, so the "score at date X" view is always reproducible from the raw data.

Stability gate (September 2026). Two research passes on the same real-world situation can disagree by a quarter point. A score change therefore lands only when the run gives evidence for it (a source URL not cited before, or a change in the named laws) or when the same change repeats on the next weekly run. Held candidates are published in pending.json. Text fields, confidence, and the last-updated date always refresh. Every recorded score change in the history therefore corresponds to new evidence, a repeated finding, or a documented recalibration.

Evidence grounding (July 2026). Research runs can be grounded in verified policy-initiative records from the OECD.AI Policy Observatory (Policy Navigator/GAIIN): for covered countries, the model scores against those records as stated facts rather than open-ended recall, and the records themselves are shown in each country's Policy initiatives panel section. Every run, grounded or not, is recorded with its model, prompt version, and token counts in a public research_runs provenance table (see Data & API).

Gold set and drift check (September 2026). Ten countries across the range of the implementation index (China, Japan, the United States, Germany, the United Kingdom, Singapore, Brazil, Kenya, Bhutan and Kiribati) have a gold score for each of their twenty sub-indicators, with a one-line justification per dimension and the sources used, published in gold_set.json. Each entry states whether it is a draft or verified: an entry is verified only when the maintainer has checked every score against the listed sources and recorded the date, and the set is re-verified each quarter. All ten entries are currently drafts awaiting that check. After every research run the scores the model returned for these countries, taken before the stability gate so the check measures the model rather than the gate, are compared with the gold scores on four metrics: the mean absolute error per dimension, the mean signed error per dimension (whether the model scores higher or lower than the gold set, which the absolute error cannot show), the share of sub-indicators within one point, and the largest single deviation. One row per run is appended to drift.json and mirrored to the gold_checks table of the API, so a model or prompt change that moves the calibration shows up as a break in that series rather than as silent drift. A run whose within-one share falls below 0.8, or that leans by more than half a point on any dimension, carries a calibration warning in its summary; it is never blocked.

Rubric versions

The scoring rubric has changed four times. Each change is a documented break in the series, never a silent drift. Earlier snapshots are kept as they were scored; nothing in the past is re-scored or offset, because past web content cannot be re-researched and an offset would be invented data.

• v1, March 2026. Integer scores per dimension, four-dimension composite.
• v2, June 2026. Four sub-indicators per dimension, quarter-point means, three-dimension composite index. A 5 meant the global frontier at the time of scoring.
• v3, September 2026. Fixed anchors. The frontier definition made scores comparable across countries but not across time: a country could fall without changing anything because the frontier moved, and weekly runs re-estimated the frontier every week. Each level now describes an observable state, as set out under Levels. The sub-indicator definitions and the output structure are unchanged. The first v3 run is also the first on Claude Opus 5, so the two changes share one break in the series, not two.
• v3.1, September 2026. The v3 anchors plus an insufficient evidence value, which is no longer scored as 1. The tie-break toward the lower of two levels now applies only when the evidence supports both. Countries with thin source coverage can move up at this break because gaps stop counting as absences.
• v3.2, October 2026. Tighter anchors. The first v3.1 run gave 5s on the strength of a single instrument, counted sub-national law and general law applied to AI, and told the model that leading jurisdictions reach 5 on most sub-indicators. A 5 now needs every element of its anchor in force, applicable and exercised (anything deferred or not yet used caps at 4); only the national level counts, with EU law counting for EU members; general law with no provision written for AI scores at most 2; a rule for one use case scores at most 3 on binding force and 2 on scope; and every implementation sub-indicator defines all five levels. Implementation scores move down at this break, most where a high score rested on one instrument, deferred obligations or state law. The one v3.1 run (28 September 2026) did not land in the dataset files, so the v3.2 break is the one that follows v2 in history.json.

Each break is recorded in history.json under breaks with its date, model, prompt version, and reason; the v3.2 break is recorded there when the first full v3.2 run lands (v3, v3.1 and v3.2 followed each other before a full run landed, so that run records a single break from v2). The timeline marks the date, and a country's score history labels changes on that date as a recalibration rather than a policy change. A relative view (a score divided by that week's maximum) is a possible display option for the frontend; it is never the stored number.

Known limitations

• Point-in-time view. Scores reflect the state of regulation at the last research pass. Fast-moving jurisdictions can change materially between updates.
• English-source bias. Non-English primary sources are under-represented, which systematically disadvantages countries whose regulatory debates happen in other languages. Claude mitigates this but does not eliminate it.
• Aggregate hides regional variance. Federations (US, India, Brazil) are scored at the national level; substantial sub-national variation is flattened.
• Draft vs. enacted. The rubric tries to distinguish these, but a bill that passes the week after a research pass will not update until the next weekly run.
• No independent audit. Scores are LLM-inferred, not human-coded, and have not been audited by domain experts. The sub-indicator breakdown in subscores.json exists precisely so that every dimension score can be checked claim-by-claim.
• Methodology changes. Sub-indicator scoring and the three-dimension composite index arrived in June 2026; fixed anchors replaced the frontier calibration in September 2026 and were tightened in October 2026 (see Rubric versions). Score movements on those dates partly reflect the methodology change, not only regulatory change.

Data access for analysts

Every data file behind this site is a static, versioned artifact: no API key, no rate limits. Load it directly into your tools:

Column definitions match this page. The git history of each file is the audit trail: every weekly update is one commit, so any past state of the dataset is retrievable and citeable by commit hash. The in-app Export button produces the same data merged into a single CSV/JSON.

For queryable access (filtered extracts, score history, the sources database, policy-initiative records, and run provenance) the same data is served by a public read-only REST API with CSV export. Endpoints and copy-paste examples live on the Data & API page.

Citing this site

Every view on the site (a single country, a comparison set, a historic date, a specific score dimension) has a stable permalink encoded in the URL query string. Include that permalink in your citation so readers can reproduce the exact view.

Suggested formats:

• APA (7th). Deane, R. (2026). AI Regulation Map [Data visualization]. Retrieved YYYY-MM-DD, from https://airegulationmap.org/…
• Chicago (author-date). Deane, Ria. 2026. "AI Regulation Map." Accessed YYYY-MM-DD. https://airegulationmap.org/…
• MLA (9th). Deane, Ria. "AI Regulation Map." AI Regulation Map, 2026, https://airegulationmap.org/…. Accessed DD Mon. YYYY.

The in-app "Cite" button on any country panel generates these strings for the current view and copies them to your clipboard.

Source code and contact

The site is open source at github.com/riadeane/airegulationmap. Issues and pull requests welcome. Made by Ria Deane.