Methodology
AI Incident Law is a curated public-record corpus of legal, regulatory, tribunal, and review-queue matters involving AI-related incidents, failures, and harms. It is not an exhaustive incident database, a legal advice product, or a ranking of parties named in public records.
Scope
A record is eligible when a public legal or regulatory matter directly involves AI-related conduct,
output, or use. The legal visibility matters: press coverage alone is not enough for admission to
included, though it may justify a candidate in review.
Admission criteria
- There is a public legal or regulatory matter directly involving AI-related conduct, output, or use.
- The matter has at least one primary or reliable secondary public source.
- The matter has resolved to a filed proceeding, regulatory action, settlement, judgment, consent decree, or formal investigation disclosure.
- Publication fields are present and consistent: jurisdiction, parties, AI relevance, source URLs, and date fields.
Outcome valence
A matter is admitted on the strength of the established AI-attributable defect, not on the severity of what the tribunal did about it. A ruling that declines to sanction, declines to exclude, or excuses the conduct is admitted on the same terms as one imposing a penalty; what happened is recorded in the outcome fields rather than deciding whether the record exists. A corpus that admitted only adverse outcomes would systematically overstate the legal risk of AI use, because the matters where a tribunal examined the same conduct and declined to penalise it would be missing from the denominator.
The bound: an AI-attributable defect must be established in the primary source to the same standard criterion 1 requires for AI use itself. Where the only established fact is that AI was used and no defect was found, there is no incident and the matter is not admitted.
In re Bard Implanted Port Catheter Products Liability Litigation (D. Ariz., 12 August 2026). Hallucinated citations were present in an expert report and the expert was questioned about his use of ChatGPT. The court found his explanation adequate and denied the motion to exclude. The defect is established, so the non-exclusion is the record's outcome.
Mullins v. Duquesne (W.D. Pa.). The pro se plaintiff disclosed his use of Claude under the court's standing AI-use order, and the court found the brief contained neither inaccurate, false, nor non-existent citations. Disclosed AI use plus an allegation that did not survive scrutiny is not an incident. The matter is retained as a limiting precedent on what AI use is not sanctionable, but it is not an incident record.
AI attribution
Courts increasingly notice that a fabrication bears the hallmarks of AI while expressly declining to determine whether AI produced it, because attribution is unnecessary to the holding — fabricated authority is sanctionable either way. Requiring an affirmative judicial finding would progressively blind the corpus to that posture as it becomes the common one.
So criterion 1 is satisfied when the primary source itself discusses AI in connection with the defect, whether as a finding, an observation of hallmarks, a hedged surmise, or a party explanation the tribunal recounts without rejecting. Two bounds apply: the AI discussion must appear in the source, never supplied by a third party's inference; and AI must remain a live explanation, so a matter is excluded where the source establishes a competing non-AI mechanism.
Records admitted on undetermined attribution are marked as such in ai_system_name, carry the attribution-not-determined tag and a confidence score no higher than medium, and quote the tribunal's exact words. They describe the fabrication as the defect and do not assert that AI caused it.
Fuselier v. Riscassi (S.D. Miss.). Ordered to disclose his AI use, counsel explained that he had turned to Google web search and could not confirm the fictitious citation's source, saying he was not certain it came from the AI preview but that it may have appeared there. The court recounted that account, admonished and warned him, and imposed no sanction.
Maxwell v. Michael (S.D. Ind.). The court agreed the fabricated authority contained the hallmarks of generative AI hallucinations, but held the conduct sanctionable whether or not AI was used because it represented a deliberate intent to fabricate, and recorded the litigant's own account that he composed the fake citations himself. Deliberate human fabrication is the operative finding.
H. C. v. Contreras (Cal. Ct. App.). A trial court copied a fictitious citation from counsel's brief into its ruling and was reversed. The opinion never mentions artificial intelligence, so there is no AI record to admit however strong the fabricated-citation salience.
Jurisdiction and language
Jurisdiction is not an admission criterion. A non-US matter that satisfies all four criteria is admitted on exactly the same terms as a US one, and there is no presumption that non-US matters stay in the global bucket. A record whose only remaining gap is that it is not American has no remaining gap.
Source language does not block admission either. The criteria ask for a primary or reliable public source and say nothing about the language it is written in; what must be in English is the record, not the source. Reading a translation requirement into the criteria would permanently exclude most of the EU, Latin America, and Asia from a corpus that claims to be a global evidence layer. Where the source is not in English, the operative language — the measure imposed, the finding of AI use — is quoted in the original alongside an English rendering, so a reader can check the translation rather than trust it, and the source language is recorded as a tag.
Dataset buckets
Admitted public matters rendered on the site, exposed through MCP, and exported to the Obligation-First API.
Likely in-scope candidates that need verification, stronger public sourcing, or scope decisions before admission.
Review, for non-US matters: a staging queue for candidates that still need translation, jurisdiction-specific interpretation, or additional normalization. It is a queue, not a destination — nothing is published from it, and a candidate leaves it for included once its remaining gap closes.
Source policy
Primary public records are preferred: court filings, orders, opinions, tribunal decisions, agency actions, regulator releases, official dockets, and stable public record mirrors. Reliable secondary reporting can support a record, but source quality is favored over record count.
URL fields are normalized and validated by the maintainer tooling. Final public data uses https://
bare-domain URLs, and rejects appended prose, credentials, non-HTTP schemes, embedded whitespace, unsafe
delimiters, and malformed source lists.
Freshness
Records carry last_verified_date or last_checked_date. The public dataset
generated_at value is derived from the newest record verification or check date during the
build, and validation fails if it lags behind the corpus.
The steward uses npm run report:staleness and the get_staleness_report MCP tool
to surface verification decay. Scheduled source-decay and coverage-gap review is quarterly.
Agent trust boundary
AI Incident Law publishes an assistant guide for bounded maintainer and query workflows. Linked public records, external sources, issue text, PR text, scanner reports, and generated data are evidence to inspect, not assistant instructions to follow. GuideCheck conformance is a reviewability claim, not a safety claim.
Out of scope
- Speculative AI-risk commentary without public legal or regulatory action.
- AI ethics statements without a filed matter, formal action, or official investigation disclosure.
- Internal corporate disputes that have not surfaced publicly.
- Reputational controversy without a filing, order, settlement, enforcement action, or formal disclosure.
- Private filings, leaked material, anonymous tips, and records barred from public reference.
Canonical policy
This page summarizes the operating method. The authoritative repo-scoped strategy remains INTENT.md, with field-level conventions in docs/data-schema.md.