Methodology

What qualifies as an incident

A specific AI system or AI-enabled tool that caused or contributed to actual harm, or a specific, named system or deployment shown or reported to create a concrete, credible risk of harm before any harm occurred. Court rulings, regulator actions and lawsuits about such harm count. The harm can be:

• security risk
• privacy impact
• physical harm
• financial harm
• misinformation
• discriminatory outcome
• operational disruption
• regulatory action
• safety failure
• significant governance failure

What does not qualify

Records that name no specific AI system and no harm are held back, even when a tracker lists them. That covers:

• policy, legislation, proposed regulation and official meetings
• warnings, forecasts and opinions about AI in general
• research and surveys about trends or jobs
• product launches, funding, contracts and business plans
• statements by officials or executives
• purely hypothetical scenarios and unsubstantiated rumors
• duplicate reporting of the same event

Source standards

Candidate events come from public incident trackers such as the OECD AI Incidents Monitor, the AI Incident Database, AIAAIC and AVID. A tracker's page is where we find an event; it never counts as evidence. Evidence comes from the original publishers: courts, regulators, government bodies, the organizations involved and the press.

Marketplace listings, social media posts and video pages are not used as evidence. A record whose sources are still being checked says so on its page and lists none until they are.

How a record is made

An automated step reads each event's source articles and writes the record from them: the facts with the source behind each one, a neutral title, a summary and body in our own words, the organizations and AI systems involved, the date, the country and the classification. It uses only what the sources state. Article text is read to write the record and is not stored.

A second automated check then answers each closed-vocabulary question, such as the category, failure modes, consequences, harm, intent, timing and industry, with a confidence. It can remove a choice no fact supports, and it flags any low-confidence answer for a person. The same check decides whether the event is in scope, and records it judges out of scope are held back. Disagreements with the tracker over the date or country are flagged too.

An editor reviews a record against its sources before their name appears on it.

Evidence status

Evidence status is derived from what a record rests on, not from its wording. The rules are applied in this order, and the first that matches sets the status:

• Confirmed: a court, a regulator or the organization itself has stated the core facts.
• Alleged: the record is itself a lawsuit, complaint or accusation that has not been decided.
• Under investigation: a regulator, court or police has opened a formal inquiry and reached no finding yet.
• Confirmed: the record cites a court or regulator source.
• Well documented: three or more independent publishers report the core facts.
• Reported: anything less.

Disputed and corrected records

Two further values are set only by an editor. Disputed means material facts are contested by a named party. Corrected means earlier reporting was materially wrong and the record reflects the correction.

Deduplication

Multiple reports of the same underlying event become one incident record. When several trackers list the same event, the record carries each tracker's identifier, and editors can move an identifier from one record to another. The old URL of a merged or renamed record redirects to the current one.

Record identifiers

Every record in the curated index is assigned a permanent first-party identifier of the form AII-1000-00007. This identifier never changes once assigned, even if the record is updated or reclassified.

Identifiers from other incident databases (OECD AI Incidents Monitor, AI Incident Database, AIAAIC, and others) are stored alongside the first-party identifier as cross-references. They allow lookups across systems but are distinct from the Index's own identifier and do not imply the record originates from or reproduces the content of those databases.

Classification

Each record receives one primary category and up to two secondary categories from the public taxonomy, plus, where the sources support them, failure modes, harm types, the people or groups harmed, consequences, whether harm occurred or was only a risk, intent, timing, industry and country.

Writing standards

Titles state established facts plainly. Words such as alleged and reportedly appear in a title only when the record is itself an undecided lawsuit, complaint or accusation, and a title says one thing caused another only when a court or official body found it.

Records about suicide or self-harm carry a note with where to find support and leave out any detail of method.

Use of AI

Each record is drafted by Raidu Incident Detection, the Index's automated system, and its page says so. A named editor reviews and approves a record once its sources are verified, and from then on the editor's name and LinkedIn link appear in its byline. A record that has not been reviewed carries no reviewer and claims no review.

Automated systems read the source material and draft each record's text and classification. Their output is checked by a second automated step and by editorial review. They do not decide whether an allegation is true: evidence status follows the published rules above, and only an editor can mark a record reviewed.

Corrections

Anyone, including the organizations named in a record, can submit a correction, response or updated source without charge. Record removal is never sold, and payment is never accepted to change an evidence status.

Updating records

New evidence can change a record’s status, classification or summary. Every edit is kept in the Index's revision log with its date, and a reviewed record shows when it was last reviewed.

Policy Monitor

Policy Monitor pages cover AI laws, regulations, guidelines, standards, strategies and governance bodies. Each entry is sourced from its official publisher, such as a legislature, regulator, government ministry or intergovernmental body, and links to the official text. Summaries are written by the Index in our own words and are not legal advice; the official text is always the authority. Each official page is checked every day, and when its status, dates or text change we record the change with its date and a link to the evidence in the policy's change history. An entry an editor has checked against its official page shows who verified it and when.

Where we can divide a policy's official text into its chapters, articles and sections, its page lists every one of them with a short note. Each note is an automated summary written from that section's text alone, in plain language and in original wording, and is checked for copying: a note that repeats eight or more consecutive words of the official text is rewritten or dropped. Long sections are summarised from their opening text, so a drafted note can miss later provisions. A note marked Drafted by the Index has not yet been read by a person; a note marked Reviewed has been checked against the official text by an editor. Every section links to its place in the official text, which remains the authority.

Each policy's controls are drafted automatically from its sections: every duty, ban, right, permission or penalty a section sets out becomes one control in a single sentence of our own words, with who it binds, when it applies and any deadline, and the same copying check as the notes. Each control is tagged with up to three of a shared set of themes, grouped in families, so policies from different places can be read against each other. A shared theme is not a legal equivalence: two policies can share a theme and still require different things. A control marked Drafted by the Index has not yet been read by a person; one marked Reviewed has been checked against the official text by an editor in our review tool. Every control links to its place in the official text, which is the authority, and none of this is legal advice.

Limitations

The AI Incident Index is not a comprehensive measure of all AI incidents worldwide. Public reporting varies significantly by country, industry, organization and incident type. Counts should therefore be interpreted as documented incidents in this dataset, not as estimates of true incident prevalence.