How to catch AI lawyers making up fake case law?

Large language models confidently cite laws that don't exist, have been repealed, or apply to the wrong jurisdiction—a critical problem in legal work. Researchers built a citation grounding metric tested against 502 million real citations from Ukrainian courts, exposing three types of hallucinations: nonexistent provisions, irrelevant ones, and outdated statutes. They also developed a training technique that uses corrupted real citations to teach models to reject bad citations with 98.5% accuracy. Code and datasets are open.