How the incident tracker is built, and how it compares with other AI hallucination databases

What counts as an incident?

A row is a court, tribunal or regulator decision in which a party's use of an AI tool in legal work is on the record: fabricated or misquoted authorities in a filing, AI-generated evidence or submissions, or a judge's own use of AI in a decision. The event date is the date of the decision or order, not the date of the filing that caused it. One row per proceeding; a later appeal, costs ruling or disciplinary outcome updates the row rather than adding a second one.

We do not record allegations without a decision, press reports of filings that no court has ruled on, or uses of AI that a court noted without criticism.

How is each row verified?

Every record links its primary source first: the judgment or order on an official system (Find Case Law, BAILII, CourtListener and the RECAP archive, CanLII, AustLII, SAFLII, court websites) or the regulator's own publication. A row whose facts rest only on press coverage is marked unverified and excluded from the headline counts until the document is read. The lastVerified date records when a person last opened the sources; the pipeline sweeps for new decisions on Tuesdays and Fridays and re-opens stale rows every 90 days.

Quotations in the linked analyses are copied from the document, with the paragraph or page. Numbers are datestamped. Corrections are appended, not silently made, and are listed on the changelog.

What does a row record?

Beyond the case name, court, date and a 40 to 60 word summary, each row carries the conduct (what was fabricated or misused), the outcome as the court imposed it (sanctions, fine, costs order, referral, warning, strike-off, suspension, dismissal, pending or other), any monetary penalty with its currency, and the AI tool only when the decision names it. Two fields are the reason the tracker exists alongside larger lists:

Actor. Who used the tool: a lawyer, a litigant in person, a prosecutor, a judge, an expert or a firm. 79 of 150 rows are coded, and the split as of 2026-09-05 is not recorded 71, lawyer 52, litigant in person 21, other 4, firm 1, judge 1.

Regulatory outcome. What the professional regulator did afterwards, separately from what the court did: a referral to the SRA, BSB, a state bar or a law society, and its disposition and date when it lands. 48 rows carry one. This is the layer that turns a list of embarrassing filings into a record of consequences, and it is why UK and Commonwealth rows are followed for months after the judgment.

How does coverage compare with Damien Charlotin's database?

Damien Charlotin's AI Hallucination Cases database is the reference list for this field and the one courts themselves cite. On 2026-09-04 it listed 2,016 decisions: 1,380 from the United States, 217 from Canada, 110 from Australia, 62 from the United Kingdom and 57 from Israel, with 1,157 involving self-represented litigants, 805 involving lawyers, 31 involving judges and 15 involving experts. It aims to be exhaustive, and it is.

This tracker is smaller by design. On the same date it held 150 rows, of which 89 are US, 10 Canadian, 6 Australian and 17 from the UK; 53 involve a lawyer or firm and 21 a litigant in person. The difference is depth, not disagreement: every row here has been read against the primary document, coded for actor and outcome, and followed into the regulatory aftermath, and 12 rows have a full analysis. We use Charlotin's list as a discovery source, cite it, and do not attempt to replicate its row count. Where the two disagree on a fact, the linked primary document decides, and we welcome corrections.

Where is the tracker thin?

By decision year the tracker holds 6 rows for 2023, 11 rows for 2024, 40 rows for 2025, 93 rows for 2026. The record before 2026 is a curated set of precedent-setting and widely searched cases rather than a census; the historical back-catalogue is being extended in verified batches, and the statistics page should be read with that in mind.

Reuse and corrections

The dataset is published under CC BY 4.0 as incidents.json. Attribute "SafeLegalAI" with a link. To report a missing case, a wrong disposition or a new regulatory outcome, use the report form; corrections are credited. The most recent row is Douglas v. Deutsche Bank National Trust Co..