- The scariest legal AI failure isn't a fake case. It's a real one cited for a holding it never made. KeyCite shows green. The citation resolves. The analysis is still wrong — and a second-year under deadline won't catch it. 🧵
- Everyone remembers Mata v. Avianca (2023): fabricated cases, $5,000 fine. Citators catch that now — a fake case resolves to nothing in the database. The industry "solved" the obvious failure. The dangerous one is still wide open.
- Stanford's peer-reviewed study (JELS, 2025): Westlaw Precision hallucinated on 33% of complex queries. Lexis+ AI, 17%. These are the purpose-built tools grounded in real databases. Not ChatGPT. The premium ones lawyers trust.
- And most of that 33% isn't invented cases. It's contextual hallucination: the dissent quoted as the majority, or a case interpreting a statute that was amended two years ago. The citator flags it green. The argument built on it still collapses.
- Concrete trap: ask for Delaware fiduciary-duty law, you get Stone v. Ritter (2006). Real case, green KeyCite. But Marchand v. Barnhill (2019) expanded that duty and later Chancery opinions narrowed how Stone applies. Good law, wrong filing.
- This isn't theoretical. Feb 2026, New Orleans: an attorney used ChatGPT AND Westlaw Precision AI — and still filed 11 fabricated or mischaracterized citations. Two tools didn't save them. Sanctioned anyway.
- Sanctions have stopped being a slap on the wrist. March 2026: the Sixth Circuit hit two attorneys for $30,000 over fabricated citations. The Charlotin database now logs 1,222 AI-hallucination cases in U.S. courts. Some top $100K.
- No single rulebook either. 300+ judges, each with their own AI standing order — some want disclosure, some certify every citation, some bar generative AI outright. Meanwhile 68% of legal pros use unapproved tools and under 20% of firms have a written policy.
- We don't replace Harvey, Lexis Protege, or your in-house model. We build the layer underneath: a citation-verification pipeline, a GraphRAG knowledge graph (14% better legal retrieval than vector RAG), and governance that enforces standing-order rules, not just documents them.
- Real question for litigators: would your current review process catch a real case cited for a holding it doesn't support — found by a second-year, the night before filing? Most can't — and that's the exposure. #LegalAI
- We wrote up the full verification-layer approach — contextual hallucination, GraphRAG, governance — here: https://veriprajna.com/solutions/legal-ai-citation-verification