A field guide to artificial intelligence's edge cases — the fines, court rulings, deepfake prosecutions, biased algorithms and viral failures that show what happens when AI meets the real world. Most evidence-grade entries link directly to a source; the rest carry a record locator and claim ceiling.
Cases carry a named authority, date, and source link or record locator, so you can cite the evidence ceiling rather than the rumour. Filter by country or theme to find local angles fast.
A structured, evidence-tiered dataset spanning regulators, courts and documented incidents across dozens of jurisdictions. Download the full JSON and cite the proof ceiling, not the hype.
The chatbot that sold a car for $1, the recruiting AI that learned to reject women, the deepfake robocall that faced a $6M fine. Real stories, plainly told, with sources and caveats.
ObscureAI is the public curated slice. LIMEN keeps the broader source-resolution conveyor visible so the gap between mined leads and safe cases is explicit.
Obscure AI is the public face of the LIMEN edge-case atlas — an attempt to catalogue where AI systems have actually gone wrong, and to do it honestly. Every entry is graded by how strong the public evidence is. We never present an allegation as a finding.
A regulator decision, court ruling, settlement or official report exists. The strongest tier.
Real, but the matter is at the charging, investigation, interim or under-appeal stage — not a final finding.
A documented vulnerability in an AI/agent system, anchored to a public CVE record.
A widely-documented incident reported by reputable outlets — fascinating and real, but not a regulator/court finding.
Want the method in full? Read the methods/data preprint (PDF) — it sets out the evidence architecture, the four denominator classes and the proof ceilings behind every count.
Spotted an error or a missing case? The dataset is open — download it, check the source links or record locators, and tell us what to fix.