The HalluCase registry has documented its 1,500th incident of AI hallucination in legal contexts — a fabricated citation in a federal court filing that cited the non-existent case "State v. Aetna, 142 F.4th 412."
The registry, launched in 2024 as an open-source CVE-style database, tracks hallucination incidents across jurisdictions, model types, and legal task categories. Each incident includes reproduction steps, the model and version used, and the jurisdiction where it occurred.
Key findings so far
- US courts account for 62% of documented incidents
- Contract analysis is the most hallucination-prone task (34%)
- Citation fabrication is the most common error type (47%)
- Fine-tuned legal models hallucinate 28% less than general-purpose models
What's next
The next release of HalluCase will include a CLI tool for automated incident reporting and a GitHub Action that runs incident checks on every PR.