AI Legal Hallucination: 1,458 Court Cases With Fake Citations
A public database tracks 1,458 court cases with AI-fabricated citations, growing 5–6 a day. We audited the data, the hallucination rates, and the escalating sanctions.
The Database Nobody Wanted to Need
In June 2023, two New York lawyers became famous for the wrong reason. They filed a brief citing six court decisions that did not exist. ChatGPT had invented them — case names, quotes, citations, all of it — and nobody checked. The case, Mata v. Avianca, ended in a $5,000 sanction and a global news cycle.
Three years later that looks quaint. A legal researcher, Damien Charlotin, now maintains a public, daily-updated database of court cases involving AI-hallucinated citations. When we ran our audit it listed 1,458 cases, roughly 915 of them in US courts, growing by five to six every single day.
We did not take that number on faith. We ran live web queries through a search API and cross-checked the results against primary reporting from Reuters, NPR, and the database itself. We wanted three things: the real scale, how often the tools actually fail, and what courts are doing about it.
How Often AI Actually Hallucinates
'Hallucination' is the polite word for an AI stating something false with complete confidence — here, inventing a case, statute, or quote that was never written. The comforting assumption is that this is a consumer-chatbot problem professional tools have fixed. The data disagrees.
A peer-reviewed study from Stanford's RegLab (Magesh et al., 2025) tested the leading paid legal research tools. Westlaw AI-Assisted Research hallucinated about 34% of the time. Lexis+ AI and Ask Practical Law AI landed around 17% — roughly one answer in six. General-purpose chatbots on legal questions ran far worse: 58% to 82% in an earlier Stanford/Yale preprint.
The Sanctions Are Escalating
For about a year, courts were lenient — a few thousand dollars and a stern footnote. That era is ending. In early 2026 a federal court in Oregon issued what is believed to be the largest penalty yet: roughly $109,700 in combined sanctions and opposing fees. US courts imposed over $145,000 in the first quarter of 2026 alone.
What It Means for How You Build
Here is the part the headlines miss: the problem was never that the model can be wrong. Every model can be wrong. The problem is shipping an answer that looks authoritative with no way to check it. A confident fake citation is worse than an honest 'I do not know.'
We argued about this internally for weeks while building HAQQ. The conclusion was boring and correct: do not ask the model to remember the law. Ask it to retrieve the law, cite the source, and surface its own uncertainty — then put a human in the loop before anything leaves the building. Every output in HAQQ carries a citation back to a primary source, and material work routes to attorney sign-off. It is less magical and far more useful.
Key Takeaways
- 1,458 documented cases of AI-fabricated citations, growing 5 to 6 per day.
- Even purpose-built tools hallucinate 17% to 34% of the time.
- Sanctions are escalating fast — record is now ~$109,700.
- The fix is architectural: grounded retrieval, verified citations, human sign-off.