The inclusion test
An incident belongs here when all three conditions are present:
- An AI system produced advice, content, a decision, or an action.
- A person or organization allowed that output to cross into a consequential setting.
- A verifiable adverse result, correction, outage, sanction, or near miss followed.
Embarrassing model output alone is not enough. Neither is a hypothetical risk with no real-world event.
The source ladder
Every brief should start with the strongest available record: court opinions, clinical reports, regulatory decisions, incident postmortems, issue trackers, correction notices, or direct statements from the involved parties.
Independent reporting is used to add context and test the primary account. Aggregators and social posts are discovery tools, not sufficient proof by themselves.
Confidence labels
- Confirmed means the central event is documented by a primary record or independently acknowledged by the involved parties.
- Documented with caveat means the central event is supported, but an important artifact—such as the original chat transcript—is unavailable.
- Developing means material facts remain in dispute or the authoritative record is incomplete.
Writing rules
Briefs separate what the AI produced, what a person did, what happened next, and what remains unknown. They avoid reconstructing prompts or motives that the record does not establish. A disciplinary referral is not called a sanction; an allegation is not called a finding; a reproduced chatbot answer is not presented as the original transcript.
The site links to source material so readers can inspect the record directly. Updates change the lastmod date and material corrections are noted on the article.
Scope and safety
Medical, legal, and financial incidents are summarized for reporting purposes, not as professional advice. Private individuals are not named unless their identity is already material to an authoritative public record.