How it works

How the site is made

whatAIryoudoing is a personal, public and non-commercial project. It has two parts: the public pages (the map, the country pages, the country table, Explore incidents, the News, Compare countries, Over time, State and AI, AI labs and this page) and a protected review area, where a person reviews the uncertain cases and can correct or undo the automatic decisions (see «Who does what» at the top of the page).

The public pages do not read the database at every visit. It works like a newspaper printed every week: after the update and the review, the published data are "printed" into static files, and visitors read that printed copy, not the archive. This is why the site is fast and can still be consulted even if the database does not respond.

The 3D map is drawn in the browser from those files: the height and colour of the columns come from the attention index, the countries in the top quarter of Opportunities by rank (the most untapped AI potential) glow. Anyone who cannot see it finds the same data in the country table.

To make heights readable from above too, the light comes in at a low angle and the base of each column is darker than its top (in the legend: "taller = more attention"); the "Tilted view" button in the legend lowers the camera to the side and shows the columns in profile, and the browser remembers the choice. From the legend you can also choose a type of AI and switch on the arcs between the developer's country and the country of the harm. On phones and low-powered computers the map is lighter: fewer columns, no opening animation, arcs off at the start.

On a country page the map becomes a "terrain" of points that rises where incidents happened, with a point for each of the 12 most recent incidents. The point sits where the sources place the incident: Claude Sonnet 5.5 reads the title and description and gives the city or region only if the text names it, only if it lies in one of the countries of the harm and only if it is sure enough (at least 70 out of 100); the coordinates come from Wikidata, only for a settlement or an administrative area of that country, and if several distant places share the same name the place is not used. The terrain includes nearby islands and land (Sicily and Sardinia, Hokkaido, Northern Ireland), not distant territories (Hawaii and Alaska, the French overseas departments); several incidents in the same place are spread around it, side by side. Incidents without a clear place stand in a row along the edge of the terrain ("place not given"), from the oldest on the left to the most recent on the right; those whose place lies outside the drawn terrain have their own group in the row ("other places"). If the countries of the harm of an incident change, its record is read again. The place does not change the index.

Clicking a point, its label or the title of an incident in the list opens the original article in a new tab (the first one under "Also reported by"; if there is none, the source's record); the "according to [source]" link stays next to it. Clicking a point without a label takes you to the incident in the list.

Where the data come from

The site uses only public sources (the full list, with licences and date of the last update, is further down, in Sources and licences):

• Incidents: AI Incident Database and AIAAIC Repository.
• AI exposure: Generative AI use in the population · Enterprises using AI (Eurostat) · Individuals using generative AI tools (Eurostat) · Businesses using AI (OECD) · AI compute capacity.
• Safeguards: Regulatory quality · Rule of law · Voice and accountability.
• Opportunities (the view and the glow on the map): Epoch AI's database of AI models (file all_ai_models.csv), the World Bank's Human Capital Index and population, developers from GitHub Innovation Graph.
• AI news (right-hand panel of the map and «In the news» in the country pages): GDELT Project, outside the index (see Daily news).
• Public uses of AI («How the state uses it» in the country pages): official transparency registers of the European Union, the United States, the United Kingdom and the Netherlands, outside the index (see Public registers).
• Companies that develop AI (the AI labs page): official signatory lists of the commitments (European Commission, gov.uk), Stanford's Foundation Model Transparency Index and searches by Claude on the official pages, outside the index (see AI labs).
• Country borders: Natural Earth, via world-atlas.
• Reports and suggestions from visitors, always after review.

The country of an incident is assigned only when the source itself gives it (the place of the harm), with one declared exception: the country inferred from the text. Incidents without a country (N/A) count in the totals but do not colour any country (in the last 24 months: 449 of 1,441).

Under each incident, "Also reported by" lists the articles already cited by the sources (the links on the AIAAIC record and the AI Incident Database reports), merged without duplicates when two records are the same incident, up to 10 per incident: only the title, outlet, date and a link to the original article, never the text. At the end there may be articles found in the news collected from GDELT: for recent incidents a program looks, among the news from a week before to a month after the incident, for similar headlines (shared words, organisations involved, country of the harm); then Claude Haiku (Anthropic) reads each pair and says whether the article reports the same incident, with a confidence from 0 to 100. From 90up the article appears marked "found in the news (GDELT), checked with AI"; from 60 to 89 a person decides (marked "found in the news (GDELT), checked by a person"); below that, or if the article reports a different event, it is discarded. A person can always undo the automatic decisions. The search runs every day and only finds news still kept (see the retention note in the daily news). They do not change the index.

AI labs

The AI labs page brings together, for 29 companies that develop AI, what they promised (the public commitments they signed), whether they published the safety framework promised in Seoul, what other people's ratings say and the incidents in which they are named as the developer. Only dated, sourced facts: no score of ours and no ranking. The page does not change the index or the map views.

• Which companies: a company is included if it meets at least one of three criteria. It signed the Seoul Frontier AI Safety Commitments (May 2024), the only commitment on the most powerful models with a precise promise that can be checked (20 companies); it is rated by the FLI AI Safety Index (9); it trained a model above 10²⁵ FLOP according to Epoch AI, the threshold above which the AI Act presumes a «systemic risk» (15). A company can meet more than one. The list is checked again by hand when a criterion changes (new signatories, a new FLI edition, new models in Epoch AI). Left out are, for example, the smaller signatories of the EU code and of the Chinese commitments, and the companies with incidents but without frontier models, which can be found in Explore incidents and Connections.
• One company, one name: in the incident sources the same company appears under different names (for example «Meta», «Meta Platforms», «Facebook»). For the companies on this page a hand-written list of names merges them into one, also in Explore incidents and Connections, where the name becomes a link to the company page.
• Where the signatures come from: every week the site reads the official signatory pages (the European Commission for the General-Purpose AI Code of Practice and for the AI Pact, gov.uk for the Seoul commitments) and keeps a copy. The site shows the signatures written in the project file: Claude researched them on the official pages and the companies' websites, with the sources cited (last search: 27 Oct 2026), following the method we decided; a person has not yet checked them all again, and will do so before the site goes live. When the official pages and the file do not match, a person checks and corrects them. The commitments without a list that can be read automatically (the 2023 White House commitments, the Hiroshima process reports to the OECD, the 2024 Chinese commitments) are not compared every week: they are checked again every three months. The official lists do not give the date of signature: where it is missing, for names that appeared after our first reading of the list the page says when we first saw them («first seen»); names already there at the first reading have no date.
• Seoul safety frameworks: the signatories promised to publish, by the Paris summit (February 2025), a safety framework on severe risks, with the thresholds above which risks are intolerable. Claude searched for the document, following the method we decided, for each company: on its websites (site maps, legal pages, blogs, PDFs, also in the company's language), on the gov.uk page of the commitments and in the Wayback Machine for the dates, without using third-party trackers; every document found was opened and read, and is linked on the company page. The result (for example «does not cover severe risks») is Claude's reading, still to be checked again by a person. Each result carries the date of the search. «Risk thresholds» only says whether the document contains them, not whether they are good.
• «No safety framework found» (today for 5 signatories) is the result of the search on the date given, not proof that the document does not exist; for some companies the website could only be read in part, and the page says so. The site never writes that a company «violated» a commitment.
• Other people's ratings: from Stanford's Foundation Model Transparency Index, licensed CC BY 4.0, we show the transparency score, with the average of the companies rated in the same edition (10 companies on this page are rated). For the FLI AI Safety Index and SaferAI there is only a link to their pages, without figures: their ratings have no open licence, and permission to show them will be asked later.
• Documents tracked over time: for each company the site tracks the main public documents, that is the safety policy, the usage policy, the training data summary required by the AI Act and, for the services archived by Open Terms Archive, the terms and the privacy policy (currently 57 documents from 23 companies). The addresses were found by Claude on the official pages, still to be checked again by a person. For 32 documents the versions come from Internet Archive's Wayback Machine: one copy a month for the history, then the latest archived copy every week. The other 25 (terms, privacy and usage policies of generative AI services) come from Open Terms Archive (GenAI Governance Archive, a database licensed ODC-By), which already cleans the texts and keeps their history; these are checked every week too.
• What counts as a change: from each copy only the text is kept (without menus, scripts and footers) and it is compared word by word with the previous version. If no word changes, it is not a new version. A change that only changes dates and version numbers, or with fewer than 20 words added and removed in total, is «minor», but only if none of the changed words is a negation («not», «never», «without»…), a word of obligation or permission («may», «must», «prohibited»…) or a number that is not a date: these can change the meaning of a sentence on their own. On the company page a minor change is a small grey point on the timeline, without a summary. The table column «Latest change to the safety policy» counts neither the first version nor minor changes or changes discarded by the reviewer.
• Summaries of the changes: for each change that is not minor, Claude Haiku receives only the changed paragraphs, with a few lines around them, never the whole document, and writes two or three neutral sentences in Italian and English on what was added, removed or rewritten, without judgements and without names of people. The summary goes online only after a person has approved or corrected it; if it is discarded, the site shows only the date of the change, with the links to the two copies. Changes not yet checked do not appear.
• Copyright: the texts of the documents belong to the companies and the site never republishes them. The full text is kept only in our private archive, to make the comparison; online there are the dates, the short summaries and the links to the company's page and to the archived copies, where anyone can read and compare the two versions.
• Conflict of interest: the site uses Claude, an AI model by Anthropic, to read and summarise texts, and Anthropic is one of the companies on this page; moreover it ranks first in several external ratings. Claude gives no judgements or rankings of the companies: the facts on the page come from the cited sources, researched by Claude and still to be checked again by a person; the categories of the Seoul safety frameworks (for example «does not cover severe risks») are Claude's reading of the documents, and each company page links the document. Anthropic's own documents are also summarised by Claude, with the same instructions and the same approval by a person as the documents of the other companies. This will also be checked again in the independent review before the site goes live.

Limits of this page:

• The official signatory lists have no dates: «first seen» can be later than the actual signature.
• Incidents per company depend on how much the press and the sources talk about it: better-known and more widely used companies have more, and more incidents do not mean less safe systems (see the limits).
• External ratings measure different things (transparency, risk management, overall safety) on different groups of companies: they cannot be compared with each other or added up into a score.
• Chinese companies publish less in English, and the 2024 Chinese commitments are read from a third-party source (AGORA, by the Emerging Technology Observatory), because the original does not open to programs: they are copied into the file, without automated collection.
• The Wayback Machine does not archive every version: a change that lasted a few days, or a page the archive did not save, can be missing, and the history starts from the first archived copy. One copy a month is enough for the history, but two changes in the same month can appear as one.
• Some pages change with the country or language of the visitor, or show different versions to different people: the archived copy can then be in another language or be a cookie page, and look like a change that did not happen. The reviewer marks these copies as «false» and they do not appear on the site.
• Open Terms Archive has tracked these services since November 2025; when a service is renamed, its history in the archive starts from the new name. It sometimes combines several pages into one text, and the company page says so.

When it is updated

Once a week, on Monday: the collection programs download the sources and save an original copy of them; the index is recomputed and saved with its date, so comparisons over time are possible; after the review the new edition of the site is "printed", that is, the pages are prepared as static files (with the command npm run esporta). Current edition: index of 11 Nov 2026.

AI news, instead, is collected every day, early in the morning, for the previous day; it reaches the site with the new edition, and the panel says up to which day it goes.

If a source does not respond, its data from the last time remain and the source table says so. The weekly rhythm and the static pages can be reconsidered if the site gets many visits or needs more frequent updates.

How we review the data

• The same incident can appear in more than one source. The system automatically proposes the pairs that are probably the same (similar title, same year ±1, organisations in common); Claude Sonnet 5.5, an AI model, pre-assesses them (same / different / uncertain, with a short reason and a confidence from 0 to 100, which is its own self-assessment, not a measured probability); then a person reviews them every week, and the final decision is always theirs. Pairs that Claude judges "the same" with a confidence of at least 90 are merged automatically, without waiting for the review; the reviewer can split them at any time, and from then on the pair stays split.
• Until a pair is reviewed, the incident counts twice. Pairs not yet reviewed that Claude judges "the same" with a confidence of at least 75% carry the label "probably the same" on the site.
• Anyone can use "Report an error" (on an incident, a match, a country figure) or "Suggest a source or a news item", without an account. Reports go into the same review queue: nothing changes on the site without a human check. Privacy notice.
• A person also reviews the texts, the methodology, the new versions of the index and the automatic classifications.
• Country inferred from the text for incidents without a country: Claude Sonnet 5.5 proposes it, a person checks it. Proposals with a confidence of at least 90 are accepted automatically (with spot checks); the reviewer decides the others.
• Signatures on the AI labs page: every week the official signatory lists are compared with the signatures in the project file (researched by Claude, still to be checked again by a person); a person looks at the differences (new or missing signatures) and corrects the file of signatures, with the date of the check.
• Headquarters of organisations: found automatically (Wikidata, then Claude Sonnet 5.5); a person decides the uncertain cases and can also correct the automatic headquarters.
• Severity: the reviewer can correct the level of each incident by hand (see the formula), also thanks to readers' suggestions.
• Summaries in the "The week" bar: every week, with the data update, Claude Sonnet 5.5 writes a short summary (in Italian and English) for each sentence, using only the site's data and their sources, without searching the web; the links to the sources come from our data, not from the model. A person reads them and approves, corrects or discards them: only the approved ones go on the site, labelled "Written with AI, checked by a person" and with their sources. Without an approved summary the sentence links straight to the news item, the country or the view.

The formulas

Attention index v0.6

Higher means more things to keep an eye on. Sum of three parts with fixed weights, each on a 0-100 scale: exposure to AI as a percentile among countries (so extreme values, like the computing capacity of the United States, do not crush the others), missing safeguards (half quality of institutions, half safeguards on data and AI) and incidents on a fixed scale (see the incidents component):

Index = 40% × incidents + 30% × exposure + 30% × missing safeguards

At least 2 real data points among exposure, safeguards and incidents (at least one) are needed, otherwise the country is "Insufficient data" (not a low score). If exposure or safeguards are missing, the part that is there counts for both and the country is "partial": it has a score but no position in the ranking. For Attention we do not show a world value: it is an average of comparisons between countries and would say little. Instead we show how many countries are in each level.

Index levels, fixed bands of 20 points: Low 0-20 · Moderate 20-40 · Significant 40-60 · High 60-80 · Very high 80-100.

The colour of the columns is continuous: every point of the index changes it a little, from dark grey (20 or less) to amber, bright orange and almost white (80 or more). The scale is fixed, so the same score always has the same colour from one week to the next. The levels are used in the texts, in the box that appears when hovering over a country and in the country page. In the view by type of AI the same scale goes from percentile 0 to 100; in the other views it goes from a fixed minimum to a fixed maximum, written for each view below.

Each score states the version of the methodology and the date of calculation; a new formula becomes a new version.

Incidents component

Sum of the weights of the country's incidents in the window, on a truly fixed scale: 10 weighted incidents (about 23 recent minor incidents) are worth 100; below, in proportion to the logarithm of the sum, so the first incidents count more than the last ones. A country's value depends only on its own incidents. An incident with more than 5 countries of harm weighs 1/N in each (N = its countries): a decision that touches 50 countries does not bring small countries to the top. Records from the same source, about the same country, in the same month and with the same main outlet count as a single episode (the highest weight): so a batch of records about the same fact does not weigh like many different incidents. A country with no recorded incidents has 0 in this part: incidents can only add, and one more incident never lowers the score. Zero incidents often means only little press coverage.

• Window of 24 calendar months: the weight drops a little every month, in a straight line: 0.45 in the current month, 0.354 11 months ago, 0.25 23 months ago; an older incident never weighs more than a more recent one.
• Each incident has a single date, chosen in this order:
  1. the exact date given by the source, or that of a confirmed twin record that has it (AI Incident Database);
  2. the month shown on the AIAAIC record page (for example "Occurred: May 2026"), even when the downloadable file gives only the year;
  3. the date of the oldest article linked to the record, shown on the site as "month estimated from the first news story";
  4. otherwise only the year. If the record page or the articles cannot be reached, we try again for at most 3 weekly updates.
  If an incident has more than one date (for example because it appears in two sources), the one found with the best method counts; if they are equal, the oldest. The weight applies to all the countries of the incident.
• Incidents with only the year, if the month cannot be reconstructed: each month of the year is worth one twelfth of its weight, and only the months inside the 24 months count. So a year that gradually leaves the window gradually loses weight, without jumps.
• Incidents with a year in the future (a mistake in the source): they weigh like those of the current year (weight 0.439).
• The weight of each incident is multiplied by that of its severity: 1 if minor or not assessable, 1.5 if of moderate severity, 2 if serious (see severity, below).

Severity of incidents (formula g1)

Each incident answers four questions; the points add up:

Trauma and serious mental-health harm count as physical or health harm (4 points); mild anxiety or distress as other concrete harm (2 points).

Total from 0 to 10: minor 0–2, moderate 3–5, serious 6–10.

Examples:

• A chatbot pushes a teenager to suicide, the family sues → 6+0+1+1 = 8, serious
• A welfare benefits algorithm discriminates against thousands of families, an authority investigates → 4+2+1+1 = 8, serious
• A deepfake video scam makes one person lose money → 2+0+0+0 = 2, minor (with arrests 3, moderate)
• A chatbot gives wrong answers without concrete harm → 0, minor

Where the answers come from:

• The answers come first from the sources: for AIAAIC the columns on harms (to people, to society, to the environment), consequences and sector (the sector approximates vulnerable people: health, mental health, education, social care, children, immigration, public welfare, justice, prisons); for AI Incident Database deaths and injuries from the CSET annotations (only older incidents) and the people affected. For both, an incident with 3 or more countries of harm affects «a very large number of people or a whole country».
• The answers that the sources do not give are inferred by Claude Sonnet 5.5 from the title and description of the incident: the label on the site then says «partly estimated from the text». The sources always take precedence and the reviewer can correct the level.
• If the type of harm cannot be understood, the incident is «not assessable» and has no label; the other questions without an answer count as 0.
• The number of articles is not used: it measures press attention, not severity.

In the index (methodology v0.6) each incident weighs 1 if minor or not assessable, 1.5 if of moderate severity, 2 if serious.

On each incident of the site anyone can say whether its severity seems minor, moderate or serious («Does it seem minor, moderate or serious?»), with an optional comment, without an account and without an email; at most one suggestion per incident every 6 days is accepted from the same address. Suggestions never change the level by themselves: the reviewer reads them and can confirm the level, correct it or give one to «not assessable» incidents. How many suggestions an incident receives does not appear on the site.

Country inferred from the text

For incidents without any country, Claude Sonnet 5.5 reads the title and description and proposes the country of the harm (at most 3) with a confidence from 0 to 100; if the event is global or the text is not enough, no country. It is an explicit exception to the rule "the country only when the source gives it": on the site these incidents carry the label "country inferred from the text", and the country given by the source always wins.

Proposals with a confidence of at least 90 are accepted automatically (with spot checks); the reviewer decides the others. In methodology v0.6 these countries also count in the index, and the country pages show them with the label.

Types of AI

Generative AI · Biometrics and surveillance · Computer vision · Autonomous vehicles and robotics · Decision-making and scoring algorithms · Recommendation and moderation · Language and speech · Unspecified AI · Algorithm without AI in the strict sense.

The rules that assign the types were checked and approved on 12 October 2026. An incident can have more than one type (for example a deepfake used for a scam with voice recognition) and counts for each; "Algorithm without AI in the strict sense" marks fixed-rule systems, so they can be kept apart.

The type comes from the labels of the sources (the "Technology" column of AIAAIC; for AI Incident Database the developer categories, e.g. "deepfake technology developers", and the CSET fields). AI Incident Database categories starting with "unknown-" (unknown developer, e.g. "unknown-llm-developer") are used only when one of the approved rules already matches them. If you choose a type, the map shows the number of incidents of that type in the last 24 months, as a percentile among the countries with at least one: it is not the attention index and does not change it. The arcs on the map are not filtered by type: they always show all incidents.

Origin and arcs

The country of origin of an incident is that of the headquarters of the organisations that developed the system; if they are not known, that of the organisations that used it, and the page says so with "from the deployer". This also applies when the developer is named but its headquarters is still uncertain: the origin can therefore change when the reviewer decides.

Where the headquarters of each organisation comes from:

1. Wikidata (open data, CC0): only a label or alternative name equal to the name counts, of an item that is an organisation (company, public body, university, association…); we take the country of the headquarters (property P159), otherwise the country (P17), otherwise the country of origin (P495). Unrelated namesakes (music bands, sports teams, record labels, fictional organisations) are discarded, but only if another organisation remains.
2. Claude Sonnet 5.5, for the names that Wikidata does not resolve and for the cases that Wikidata leaves uncertain (in that case it also reads Wikidata's candidate countries, besides the titles of the incidents in which the name appears): it proposes the headquarters with a confidence from 0 to 100, and the proposal counts only with a confidence of at least 80. If Wikidata left the case uncertain and Claude answers, even confidently, that the name is a person or a generic category, the case stays uncertain (it contradicts Wikidata, which found organisations): the reviewer decides.
3. Reviewer's correction: a person decides the uncertain cases (several organisations with the same name in different countries, headquarters in more than one country, low confidence); until they do, those incidents stay without an origin. The reviewer can also correct headquarters found automatically, and their corrections are never overwritten. Natural persons and generic categories (for example "unknown scammers") never have an origin.

The arcs on the map go from the country of the developer's headquarters to the country of the harm, when they differ: we show the 20 links with the most incidents in the last 24 months between countries drawn on the map, brighter where there are more incidents. They are switched on and off from the legend. Clicking an arc (or the list of arcs in the legend, with the keyboard) opens a panel with the number of incidents, the organisations based in the country of origin and the list of incidents, each linking to the card of the country of the harm.

Index of previous years

For the years before the launch the index is reconstructed backwards, at 31 December of each year, with the data available today: some sources keep only the latest value, so in distant years there are more countries with "insufficient data". Reconstructed points are always marked as such and do not enter the weekly comparisons.

AI exposure and missing safeguards

For each indicator we use the country's most recent value from the last 3 years, including data from the current month. Each indicator becomes a percentile among the countries that have it; the component is the average of the indicators available for the country. The percentiles of the Eurostat indicators are computed among EU countries only and then averaged with those of the other indicators, computed over the whole world.

Missing safeguards = 100 − average of (average of the percentiles of the quality of institutions; average of the safeguards on data and AI on a fixed scale)

More safeguards, less attention: this is why the average is inverted.

From version v0.3 of the index, from V-Dem (Digital Society Project) we use two measures on online privacy laws. We deliberately left out the «government capacity to regulate online content», because a high value can also mean censorship.

The map views

The map can be read in five ways. It opens on the attention index; the buttons above the map change the view and each view has its own colour.

• Attention (how much attention AI use deserves in each country): the attention index (see below). Without data: «insufficient data».
• Adoption (how much people and businesses in the country use AI): people, businesses and developers who use AI, on an absolute scale from 0 to 100: it no longer coincides with the "AI exposure" component of the index. Without data: «data not available».
• Regulation (how many rules and safeguards the country has for AI and data): rules that exist, rules that are applied and institutions for AI, on an absolute scale from 0 to 100: it no longer coincides with the "safeguards" component of the index. Without data: «data not available».
• Problems and attacks (reported AI incidents (24 months), online manipulation, digital abuse and surveillance in the country): AI incidents per million AI users (0 without incidents), foreign online manipulation, state digital abuse and AI surveillance, on an absolute scale from 0 to 100. The AI type filter is here, on incidents only. Without data: «data not available».
• Opportunities (where the country is ready for AI but still uses it little): Readiness × (100 − Adoption) / 100: untapped potential (see below). It does not enter the index; in the Attention view the top quarter of countries by rank glows. Without data: «data not available».

Adoption, Regulation, Problems and Opportunities are numbers from 0 to 100 on an absolute scale (views configuration v1): each measure goes from 0 to 100 with fixed limits, never recomputed on the data of the week, so the world number moves when the world changes. The attention index still uses percentiles among countries for exposure and missing safeguards.

Adoption and Regulation no longer coincide with the «AI exposure» and «missing safeguards» components of the index: they also use other sources and other scales. In the country table, clicking a value, «How we get to N» shows the calculation of each view.

How it is computed: each measure goes to 0-100 according to its scale; within a group (or an item) the measures of the same organisation are averaged first, then the organisations are averaged (one source, one vote); the groups are added with their weights. Only the final number is rounded, to one decimal.

For each measure we use the country's latest value from the last 3 years (GovTech Maturity Index, Research and development expenditure 5 years; AI surveillance (Carnegie, 2022), Human Capital Index 10 years).

Adoption checks: «How we get to N» also shows the country's position in two rankings published by companies on the use of their own services, OpenAI Signals (ChatGPT messages per inhabitant, personal adult accounts only, without business use) and the Anthropic Economic Index (Claude use per working-age inhabitant), next to its position according to Microsoft, the source of the formula. Positions are counted among the countries present in every ranking, each with its most recent period.

The checks do not enter the formula. They are relative: they only say who is above and who is below (OpenAI publishes only the position, Anthropic the comparison with the world average, which is always 1), so they cannot say whether AI use is growing in the world. Each company sees only its own users and its own products. And those who sell AI must not decide the site's numbers: it would be a conflict of interest. Some countries are missing: ChatGPT is not available for example in China, Russia, Iran, Belarus, Venezuela, Cuba, Syria, Hong Kong and North Korea, and the Anthropic Economic Index covers about 120 countries.

When the position according to the formula is further than 33% of the common countries (a third of the ranking) from that of both checks, the warning «The companies' checks do not agree with the formula: read the number with caution» appears. The number does not change.

Sources of the checks: OpenAI Signals v2.0 — openai.com/signals (CC BY 4.0); Anthropic Economic Index — huggingface.co/datasets/Anthropic/EconomicIndex (CC BY).

Licence notices: Contains information from the Stack Overflow Developer Survey, made available under the ODbL — survey.stackoverflow.co (contents under DbCL 1.0). The per-country values published here are also under the ODbL 1.0. Contains public sector information licensed under the Open Government Licence v3.0. UK Government, members of the International Network for Advanced AI Measurement, Evaluation and Science (press release of 9 December 2025) — gov.uk.

Opportunities: untapped potential

Opportunities = Readiness × (100 − Adoption) / 100, without a percentile: a number from 0 to 100 on an absolute scale (Readiness and Adoption: see the views above). It is high only where the country is ready for AI and still uses it little; it is low both where the opportunity has already been taken (ready and using it a lot) and where the foundations are missing (not very ready and using it little).

For example, a country with Readiness 80 and Adoption 20 has Opportunities 80 × (100 − 20) / 100 = 64; with Adoption 90 the same country would have 8.

It is untapped potential according to the available data, not an economic forecast: it shows where the foundations are there and use is still low, not how much the country will grow. Opportunities stay outside the index. In the Attention view the countries in the top quarter of Opportunities by rank (among countries with a value) glow on the map, more intensely towards the top: the first country at full strength, the last of the quarter at one fifth. It can be switched off from the map legend.

Limits to keep in mind:

• In China and Russia Microsoft measures low adoption partly because many foreign AI services are not available there: their Opportunities may come out higher than they really are.
• The «Research and AI» group is missing for many countries until a source with the research of almost every country arrives (OpenAlex): there it counts as 0 and lowers Readiness.

Daily news

The right-hand panel of the map shows the AI news of the last 7 days, chosen for the map view; a country page shows the news that mentions it («In the news») and how many AI articles mentioned it each day in the last 30 days. News does not enter the index or the views.

• Source: the Global Knowledge Graph of the GDELT Project, which reads online outlets from all over the world every quarter of an hour, in English and translated into English. We keep only the headline, outlet, date, link and the data GDELT extracts automatically (places mentioned, tone), never the text: the headline links to the outlet's article. The outlet is assigned to a country with GDELT's domain table, with a few manual corrections.
• Filter: a news item is about AI if the headline (or the address) contains words such as «artificial intelligence», «AI», «ChatGPT», «deepfake», in these languages: English, Italian, Spanish, French, German, Portuguese, Chinese, Russian, Arabic, Japanese, Korean. Acronyms (AI, IA, KI) count only in capitals, so they are not mistaken for common words (in Italian «ai» is a preposition).
• Stories: articles with the same headline within 3 days are one story; headline and link are those of the first article, and «and N other outlets» says how many other outlets published it. Stories are ordered by number of outlets, then most recent first (at most 30 per view and 10 per country).
• Provisional automatic classification, with rules and keywords per language (later Claude will do it): problems (deepfakes, scams, attacks, data breaches, discrimination, lawsuits and fines, surveillance); safeguards (laws, regulations, AI Act, authorities, guidelines, bans); benefits and new uses (health, schools, agriculture, climate, research, new models); adoption (businesses and public bodies using AI, users, investment); economy (stock market, valuations, listings): not shown; general (no signal). A news item can have more than one category. It is rough and can be wrong.
• Which news in which view: Attention: problems, safeguards, benefits and new uses, adoption, general; Adoption: adoption, benefits and new uses; Regulation: safeguards; Problems and attacks: problems; Opportunities: benefits and new uses.
• The news of the week (since 26 October 2026, replacing «Worth watching this week» chosen by a person): at the top of the panel, the 10 stories covered by the most outlets since Monday («All» tab) and the 10 about problems («Watch list» tab), with identical stories merged and classified automatically by Claude; a person can remove a story from the list. Each week the list is kept, and the News page shows it at the top of that week.
• Retention: the full list of articles is kept for 10 days; stories published by at least 2 outlets or classified «problems» are kept for 60 days (headline and data only, without the articles), because they are needed to find the news about an incident recorded weeks later; daily counts per country and category, the stories chosen for «Worth watching», the articles confirmed under an incident and the stories picked up by at least 5 outlets are kept forever, the latter with their headline, link and the links of at most 6 different outlets (including the chosen one). If collection stops for more than 7 days, the panel goes back to the latest recorded incidents.

Public registers: how the state uses it

A country page shows, when they exist, the AI systems of public administrations described in official transparency registers: how many are in use, piloted, in development or retired, the main sectors and the most recent cases, with a link to the register. A register shows how transparent the country is about its systems, not how much it uses them: a country with a well-kept register would look «further ahead» just because of that. This is why registers never enter the map, the index, the views or the countries table, and countries without a register have no such section (no «zero»).

• The European Commission's Public Sector Tech Watch (CC BY 4.0): use cases of public administrations in Europe, only those with AI as the primary technology; cases of the European Union as a whole do not belong to a country and are left out.
• The United States' Federal AI Use Case Inventory, 2025 (public domain): for the United States the main sectors are the federal agencies.
• The United Kingdom's Algorithmic Transparency Recording Standard (Open Government Licence v3.0).
• The Netherlands' Algoritmeregister (reuse permitted by the site): AI systems only. The register has no «type» field, so an algorithm counts as an AI system if it is classified as a «high-risk AI system» or if its description mentions AI, machine learning, language models and the like (without negations such as «geen AI»). It is an automatic rule and can be wrong.
• The Government of Canada's GC AI Register (Open Government Licence – Canada): AI systems of the federal government; the vendor and personal-information fields of the file are never read.
• Catalonia's Registre de sistemes d'intel·ligència artificial (Open information use licence – Catalonia): a regional register, so in the card of Spain it is never the main register and is mentioned next to Public Sector Tech Watch.
• The ODAP inventory of public algorithms (Observatoire des algorithmes publics, CC0 1.0) for France: it is kept by a civil society association, not by the state, so it is mentioned next to Public Sector Tech Watch; machine-learning algorithms only.
• One register per country: the national one, when there is one, before Public Sector Tech Watch; other registers with cases from the same country are mentioned with their number, without adding them up, because they may describe the same systems.
• No personal data: we keep only title, body, sector, status, year, link and a short description from the source (at most 300 characters), from which we remove email addresses and phone numbers; the contact columns are never read. Titles and descriptions stay in the register's language.
• Updates: once a week with the other sources; a register that has not changed is not rewritten.

The state and AI

Two readings of the state for each country, from 0 to 100 on an absolute scale like the map views (configuration v1): Openness to AI and Exposure to threats. Exposure has two parts, always also shown separately: the state as a target (reported cyber attacks against institutions and foreign disinformation) and the state as a cause (social media monitoring, government disinformation, AI surveillance, spyware, internet shutdowns). A country heavily attacked from abroad and a country that watches its own citizens must not be confused.

They are computed like the map views: each measure goes to 0-100 according to its scale; within a group one source, one vote; the groups are added with their weights; only the final number is rounded. For each measure we use the country's latest value within its maximum age.

Exposure to threats = 50% × as a target + 50% × as a cause. Both parts are needed: with only one, exposure is «Insufficient data», and the part that exists is shown anyway.

Four quadrants: openness on the horizontal axis, exposure on the vertical axis, with the boundary fixed at 50 out of 100 (a value equal to 50 counts as high): «Open, little exposed», «Open and exposed», «Closed, little exposed», «Closed and exposed». The boundary is our choice, the same for both readings and for all countries: if almost all countries sit on one side, it shows. Without openness or without exposure a country has no quadrant. On the page the vertical axis can also show just one of the two parts. The State and AI page shows the four quadrants, a map with their colours and the table of all countries.

What protects: Government access to personal data (V-Dem, expert scale) (V-Dem expert scale from 0 to 4, × 25); Rule of law (0-100 score) (score already from 0 to 100); Data protection law (0 = no, 1 = partly (draft, signed, pending), 2 = yes: 0, 50 or 100); Rules on AI used by the state (GIRAI) (score already from 0 to 100), with the scales of the Regulation view and the latest value of the last 3 years. They are shown next to the exposure and do not lower it: subtracting them would mix this reading with Regulation.

Not part of the attention index, the views or the home map: these readings describe the state, not how much there is to watch about AI in a country, and they use different sources, some of them old. They appear in the country card, on the State and AI page and in Compare, and they change no other number on the site.

Limits to keep in mind:

• Little AI in the data: openness has an AI-specific measure (Oxford Insights); exposure mostly measures the digital sphere of the state (cyber attacks, surveillance, repression), where AI amplifies risks that already exist. There is not yet a worldwide measure of threats made with AI against states.
• Reports, not measured facts: EuRepoC attacks count more where the press is free and reports them. This is why the scale is logarithmic and we write «reported».
• Ceiling of the scales: from 100 attacks and from 100 internet shutdowns upwards the score is 100. Only severe cases are at the top (the United States for attacks; India and Myanmar for shutdowns): telling them apart would not change the picture, and a higher ceiling would lower all the other countries.
• Internet shutdowns: those ordered by non-government actors do not count (for example Russian forces in the occupied areas of Ukraine). Those with an «unknown» orderer (about 1 in 7 over the last 5 years, mostly in Iran, India and Pakistan, mainly during protests) count as ordered by the authorities: in practice they almost always are, even if nobody said so.
• Old sources: AI surveillance (Carnegie, 2022) 2022, Commercial spyware and forensics vendors used by the state (Carnegie, 2023) 2023; EuRepoC goes up to 2024.
• Assessments by experts and by governments themselves: Oxford Insights and V-Dem are not direct measures.
• New indices: Oxford Insights changed its method in 2025, so comparisons over time are valid from then on.
• Overlapping measures: four of the nine indicators of the Oxford Insights pillar come from the World Bank GovTech Maturity Index, which is also the «Digital foundations of the state» group. The two measures almost always move together (a very digital state usually also has a high Oxford score): so as not to count the same thing twice, GovTech weighs only 15%.

Sources and licences: Oxford Insights — Government AI Readiness Index (CC BY-SA 4.0); World Bank (CC BY 4.0 (UNESCO UIS data: CC BY-SA 4.0)); EuRepoC — European Repository of Cyber Incidents (CC BY-NC 4.0 (non-commercial)); V-Dem — Digital Society Project (CC BY-SA 4.0); Carnegie — AI Global Surveillance Index (CC BY 4.0); Carnegie — Global Inventory of Commercial Spyware & Digital Forensics (CC BY 4.0); Access Now — #KeepItOn (CC BY 4.0); UNCTAD — Global Cyberlaw Tracker (Under review); GIRAI — Global Index on Responsible AI (CC BY 4.0); EuRepoC: non-commercial licence (CC BY-NC 4.0).; Oxford Insights: CC BY-SA 4.0 licence, so its values republished here stay under the same licence.

Examples

Examples, not a count: the card section shows up to 5 AI tenders and up to 5 incidents where the state uses AI, and for Italy the decisions of the authorities and the context. They are not counted or compared between countries: how many tenders or incidents we find depends on how much a country publishes, on languages and on sources, not on how much the state uses AI. A country with examples but no readings has the section with «Insufficient data» and the examples.

AI tenders (TED): notices from TED — Tenders Electronic Daily (Reuse with attribution (Decision 2011/833/EU)), the journal of European public procurement. So only European Union countries (and the few other European countries that publish there) and only contracts above the EU thresholds: smaller ones do not reach TED. We keep calls for tenders, contract awards and direct awards from the last 24 months (no prior information notices) only if the procedure title contains a strong AI term in one of the EU languages («artificial intelligence», «machine learning», «chatbot», «facial recognition»…), never the acronyms alone: «AI» in Italian is also a preposition. In the test of 26 October 2026 about 27 notices out of 30 were really about AI. The most recent ones, one per title, with the title and the buyer in the notice's language. If the buyer is a natural person or may be a private one (bodies subsidised by an authority or with special rights, according to the legal type written in the notice), neither the buyer nor the title is shown: only «AI tender notice», the date and the link.

Incidents where the state uses AI: among the card's incidents (last 24 months), those that AIAAIC places in a government sector («Govt», for example police, welfare, immigration) or where the system is used by a body with a public-sector name (police, ministry, government, city, council, parliament, court, armed forces, border police, national health service…), matched as a whole word. Politics is not enough: a political deepfake or chatbot is not the state using AI. It is an automatic rule and can be wrong. The latest 5, most recent first; the link leads to the incident in the card's list.

Italy box: decisions of the authorities (for now of the Italian data protection authority, the Garante) on AI in public bodies and the legal context, chosen and checked by hand on the official pages, with one line written by us and the link to the source. Never names of natural persons.

Media coverage

One label per country, high, medium or low: how many AI articles from outlets all over the world collected by GDELT cite the country in the last 90 collected days (see «News of the day»). It is context for reading the other numbers, not a judgement on the country: not part of the index or the views.

Fixed thresholds (configuration v1): high from 1,000 articles in 90 days, medium from 100, low below 100. While fewer than 90 days are collected we count those we have, write «in the last N days» and the thresholds go down in proportion (with 45 days: high from 500 articles, medium from 50). With fewer than 14 collected days no label is given.

Where it appears: in the country card and in the «The state and AI» section, in the country table and on the State and AI page. When it is low, next to the country's incidents there is the note «Low media coverage: few recorded incidents may also mean little press coverage».

Limits to keep in mind:

• GDELT mostly reads online outlets, in some languages.
• A country can be cited a lot because of a single big event.
• The cited country is recognised automatically and can be wrong.

Explore incidents

The Explore incidents page lists the incidents of the last 24 months, the same ones that count in the index, from AIAAIC and AI Incident Database. An incident appears only once even if it touches several countries, and incidents merged as duplicates appear once, as in the country pages. It opens on the last 3 months. The data are those of the latest export: the page filters in the browser and does not read the database.

• Filters: search in the title (ignoring capitals and accents), country or area (a country, the European Union or a continent according to the UN M49 classification: an incident is included if at least one of the countries of harm is in the area), period, severity, type of AI, organisation and source. The filters are kept in the page address: a search can be shared with a link. Incidents with only a year have no month and appear only under «All 24 months».
• «Who is involved most» summary: the top 5 countries of harm, the top 5 organisations (with the country of their headquarters, when recognised) and the top 5 types of AI, counted only on the incidents shown by the filters. An incident with several countries or organisations counts once for each, so the bars do not add up to the total. Clicking a bar adds that filter.
• Organisations: those named by the sources as developers or deployers of the AI. AI Incident Database writes many actors as codes, in lower case and with hyphens (for example «adolescent-boys»): of these, the column, the filter and the summary show only the recognised organisations, with a known headquarters, written in readable form; the others, often generic groups, stay in the sources. The site never names natural persons among those who develop or use AI (here, in the origin of incidents and in the arcs of the map): incidents are attributed only to organisations, and the names of people mentioned by the sources stay only in the sources' own titles and links.
• Table: 50 rows at a time, sortable by date, country, severity and organisation; clicking a row opens the details, the same as in the country page.

Limits of this page:

• Incidents without a country (449 of 1,441 in the last 24 months): the Country column shows «not given by the sources» and they fall in no country or area. They are mostly the most recent incidents, which come largely from AI Incident Database, which does not give the country; the sources sometimes add it later. In methodology v0.6 the countries inferred from the text also count.
• An incident appears only if the press reported it and one of the two sources recorded it: few incidents do not mean few problems (see the limits).
• Until the pairs of duplicates are reviewed, the same event recorded by both sources may appear twice.

Connections

The Connections page shows where the AI involved in incidents comes from and where the harm happens, in a flow diagram: on the left who develops the system, on the right the country of the harm; the thicker the band, the more incidents. It uses the same incidents as Explore incidents (the last 24 months), from the latest export: the flows are computed in the browser and do not change the index or the views of the map.

• What counts as a connection: an incident links each organisation that developed the system, with a recognised headquarters (see Origin and arcs), to each country of the harm. In the «Companies → countries» view the developers are on the left, in the «Countries → countries» view the countries of their headquarters. An incident with several developers or several countries of harm counts once for each, so the bands do not add up to the total; within a band it counts only once. Natural persons never appear.
• The diagram shows at most 12 origins and 10 countries of harm, the most frequent ones with the chosen filters; the others are added up in «Other companies» (or «Other countries») and «Other countries». Filters: period (last 6, 12, 24 months), type of AI, severity and «Only between different countries». The choices are kept in the address (?vista=paesi&sel=US>GB): a connection can be shared with a link. Clicking a band lists its incidents alongside, each linking to the card of the country of the harm; clicking a name shows all its connections. The table of the connections opens with «See as a table»; on phones, instead of the diagram, there is the list of connections, most frequent first.
• Coverage: not all incidents have a known developer and headquarters, and not all have a country of harm (in the last 24 months 801 of 1,441 incidents have a known developer and headquarters, and 550 also have the country of the harm and are in the diagram). The page shows the coverage with the chosen filters at the top.
• Cases within the same country: by default they count (for example a developer based in the United States and the harm in the United States); 328 incidents are in the diagram only because of such cases. The «Only between different countries» box removes them. The arcs of the map, instead, show only pairs of different countries; from the panel of an arc, «Explore in Connections» opens this page with the same pair of countries.

Limits of this page:

• The headquarters is the developer's, not the country where the system was used or sold: a connection does not say who brought the system to the country of the harm.
• Incidents without a known developer, with a headquarters still uncertain or without a country of harm stay out of the diagram.
• The same company may appear under different names, as the sources write them (for example with and without «AI» or «Inc.» in the name, or under the name of one of its divisions). For the companies on the AI labs page the names are merged by hand; for the others, the connections of the same company may stay split into several bands.
• Bigger and more widely used companies are mentioned more often by the press and the sources: more incidents do not mean less safe systems (see the limits).

The News page

The News page shows the stories about AI of the last 7 days (up to the last day collected) picked up by at least 2 different outlets, with the same collection and classification as the map panel (see Daily news); news only about the economy stays out, as in the panel. At the bottom, «Previous weeks» loads, one week at a time, the stories kept for up to 60 days, with the same threshold; older weeks, grouped by month, only have the stories picked up by at least 5 outlets, the only ones kept forever (the page says so above the list). When a previous week is chosen, its news of the week appear at the top (the 10 with the most outlets, kept even after the stories are deleted).

• Filters: search in the title, period, lens (the map view), country mentioned, country of the outlet and language; these too are kept in the address. Coming from the map, the active view is kept; from the menu, all views are shown.
• Summary: the top 5 most mentioned countries and how many stories fall into each view; a story that mentions several countries or fits several views counts once for each. «General» are the stories about AI that fit no thematic view.
• Story details: the chosen outlet and its country; the other outlets with their link (at most 15) as long as the articles are kept, that is 10 days, then only their number (for stories picked up by at least 5 outlets the links of at most 6 outlets remain, including the chosen one); the average tone computed by GDELT, in words: negative below -2, positive above 2, otherwise neutral.

Limits of this page:

• Titles are in the original language: GDELT does not translate them (Claude may do it later).
• The classification into views is automatic and provisional: rules on the words of the title, which can be wrong.
• The same news item may appear several times with different titles: stories are merged only when the title is the same.
• The English-language press is over-represented, that of countries with few outlets read by GDELT is under-represented.
• The news does not change the index or the map views.

Compare countries

The Compare page puts 2 to 4 countries side by side, picked from the field at the top or from the links on the country page and in the country table, with the World in a fixed column. The countries are kept in the address (?paesi=IT,FR,DE): a comparison can be shared with a link. It does not read the database: it uses the data of the last export, the same as the map and the country pages.

• What is compared: the five views of the map (Attention, Adoption, Regulation, Problems and attacks, Opportunities), each followed by its parts, with the weight of the part in brackets. For Attention the parts are the three components of the index (incidents, exposure, safeguards); for the others they are the groups of the view, the same as in «The map views». The number of a view has one decimal, that of a part is a whole number. The colour of the numbers is that of the view.
• The underline: in each row it is the highest value among the countries (the World does not count; on a tie all are underlined). It does not mean «the best»: for Problems and attacks a higher value means more problems. The comparison is made on the numbers as they are written (with 1 decimal for the views), so two numbers that read the same are underlined together. With a single value in the row nothing is underlined.
• The World has only the value of the views, not that of the parts. The world value of a view is computed on the whole world, weighted by the population of the countries (for incidents the ratio of totals), and then combined with the same group weights: see «The map views». It is not the simple mean of the countries in the table and its parts are not published, so in the rows of the parts the World column is empty.
• A click on the name of a view opens, for each country, «How we get to N»: the measures with the raw values and the sources, the same detail as in the country table. «—» means data not available (or insufficient data for the view); a group marked «absent» counts as 0 and keeps its weight; «single source» and the warning about disagreeing Adoption checks appear as in the country table.
• Below the views, the two rows of «The state and AI»: openness to AI and exposure to threats, from the country file. They are not part of the index: no underline and no World. «—» means insufficient data (for openness, also a country that Oxford Insights does not cover) or a country without these readings.

Where the concrete data under the table come from:

• Incidents in the last 24 months and AI models released in the last 12 months: the same numbers as in the country table and on the map. The World is the total (incidents: 1,441, of which 449 without a country; models: 283).
• AI Act: «yes» or «no» (the European regulation applies only in the countries of the European Union; for the other countries it is «no»).
• AI authority or institute: «AI Act authority» (named by the European Commission for the AI Act), «authority awaiting appointment» (the EU country has not yet named the authority), «member of the international network» (the country's AI safety institute is in the international network of institutes), «none». «—» means no data: it happens outside the European Union and the network, where the source says nothing, so we do not write «none».
• Council of Europe AI Convention (CETS 225): «ratified», «signed» or «no», from a hand-written list (config/letture/fatti.json, with the date of the last check). «ratified via the EU» means the value comes from the European Union's ratification and not from the country's own: it is a provisional choice, to be re-checked before the site goes live.
• Cases in public AI registers: the cases of the country's main register, with its name below (other registers are not added up: see Public registers); «no register» if the country has none among those we read.
• AI news citing the country in the last 30 days: the sum of the daily articles in the GDELT collection (see Daily news); «—» if the news are missing. It is not a measure of how much the country uses AI: it only says how much it is talked about, in a press where English is over-represented.

Limits of this page:

• A comparison puts together numbers of different quality: countries with few sources have more uncertain views (see «single source» and the limits).
• The concrete data do not enter the index or the views: they help understand, not compute.
• If a country's file cannot be downloaded, its column shows the views and «details not available» in place of the parts.

Over time

The Over time page shows how the map views, the incidents and the news change, for the World or for one country. Each series has its own pace and says so on the page. Like the other pages, it does not read the database: it uses the data of the last export. The choice, the view and the two dates of «Two dates compared» are in the address (?paese=IT, ?lens=problems): a comparison can be shared with a link.

• Views: one point per week, that is the last computation of the index in that week (Monday to Sunday); if the index is computed more than once in the same week, only the last one counts. Each point carries the views saved on the same day. Changes are measured on the last computation: the «week» against the previous computation, the «month» against 4 weeks before, the «year» against 52 weeks before. Where there are not enough points the page says «not yet» and tells how many are needed.
• Reconstructed years: only for Attention we also have the years 2023, 2024 and 2025, computed backwards with today's data (see Index of previous years); they are marked «reconstructed» and have filled points, while weeks have empty points. Until there are 52 weekly points, the «year» change compares the last two reconstructed years and says so («reconstructed years»): the same holds in the table of biggest changes. The other four views exist only since 14 October 2026, when they were computed for the first time: before that there is nothing to show.
• Source changes: sometimes numbers jump because a new source arrives or a formula changes, not because the world changed. These changes are written by hand in a list (config/tempo/eventi.json, with the date of the first computation with the novelty and the views concerned) and on the charts they appear as a mark on the axis, with the text on mouse hover. A change that crosses one carries the note «includes a change in the sources», and in the two dates compared a note appears if there is one between the dates. Today's: none.
• Index version change: if the site's index version changes (for example from v0.1 to v0.2), the series restarts from the first week of the new version; the points of earlier versions stay visible in grey, with the version name, but are not compared with the new ones.
• Incidents per month: bars from 2016 to today, with the same date the index gives each incident (see Incidents component) and the same country of harm; the World counts all incidents, including those without a country. Incidents with only the year have no month: they are not in the bars and are counted apart («plus N incidents with only the year»). The last 6 months, including the current one, are dashed and marked «incomplete»: sources record many incidents weeks or months later, so those months are still growing. Do not read the bars as «incidents are falling»: recent months are always lower than the truth.
• News per day: the AI articles of each day that cite the country, from the first collected day (10 October 2026). They are a volume of articles, not of stories: a piece picked up by many outlets counts many times. With a view other than Attention only the categories of that view are counted, the same as in the news panel (see Daily news): Attention: all the news about AI; Adoption: adoption, benefits and new uses; Regulation: safeguards; Problems and attacks: problems; Opportunities: benefits and new uses. The World is the total of the day's AI articles, counted by the outlet's country; for about 1% of articles the outlet's country is unknown and is missing from the total. The comparison on the page is «last 7 days against the 7 before», and needs at least twice as many collected days.
• Biggest changes in the world: for the chosen view, the 5 countries that rise the most and the 5 that fall the most in the chosen period; only countries with a value on both dates count.

Limits of this page:

• The series are young: at first there are few points, and «year» or «month» changes arrive only with time.
• A change does not say why: it can be a real change, a new source, data that arrived late or a correction by the reviewers.
• The news say how much a country is talked about, not how much AI is used, and English-language press is over-represented.
• The series do not change the index or the map views.

Sources and licences

The site's data are published under the CC BY-SA 4.0 licence, because they derive from sources with this licence (AI Incident Database, AIAAIC; AIAAIC allows only non-commercial use). The licence of the UNCTAD data is under review and will be clarified before the site goes live. Country borders: Natural Earth (public domain) via world-atlas (ISC licence).

Limits

• The number of incidents also depends on how much the press covers each country: few incidents do not mean few problems. What we cannot see: incidents never reported by the press, reports sent to the authorities and not published, uses of AI in countries with little press coverage.
• Incidents without a country (N/A) count only in the totals: in the last 24 months there are 449 of 1,441.
• AI Incident Database does not give the country of incidents from 2025 and 2026: for recent years almost all countries come from AIAAIC.
• AIAAIC sometimes gives many countries for a single incident: in that case the incident counts for each of them.
• Until the duplicate pairs are reviewed, the same event present in two sources counts twice.
• Some territories have no columns of their own on the map, because they are too small for the grid or outside the drawn area; they are in the country table: Andorra, Antarctica, Bonaire, Sint Eustatius and Saba, Bouvet Island, Christmas Island, Cocos Islands, Gibraltar, Liechtenstein, Svalbard and Jan Mayen, Tokelau, Tuvalu, United States Minor Outlying Islands.
• The French overseas departments (French Guiana, Guadeloupe, Martinique, Mayotte, Reunion) are drawn inside France, as in the world-atlas borders; their data remain in the country table.
• Exposure is based mainly on Microsoft data on the use of generative AI; the Eurostat indicators cover only Europe (see the formulas).
• Severity partly estimated from the text, the country inferred from the text and the headquarters of organisations found automatically can be wrong: errors can be reported with the «Report an error» button.
• Comparisons over time: many incidents have only the year or an estimated month and the indicators are annual or quarterly, so finer comparisons make no sense. The weekly comparison applies to the index from the first publication onwards.
• News: the English-language press and that of countries with many online outlets are over-represented; China and other countries with few outlets read by GDELT, or with a controlled press, are under-represented. The countries mentioned are extracted automatically by GDELT and can be wrong (a city name shared by two countries, a place mentioned in passing); the AI filter and the classification are automatic and provisional too.
• The site is in Italian and English; titles and descriptions of incidents stay in the language of the source (almost always English).