Who Checks the Citations? Benchmarking Legal Hallucination Detection

With significant uptake by attorneys, judges, and the public, AI is reshaping legal practice. But hallucinated legal citations pose a growing challenge. Despite widespread expectations that newer models would improve the status quo, hallucinations remain widespread. We identify over 900 filings containing fabricated citations across U.S. jurisdictions. Courts have imposed escalating sanctions and pointed to compounding verification burdens on judges and counsel. We assess whether AI can ease this burden by automatically checking for hallucinations and citation errors. We introduce a taxonomy of legal citation hallucinations grounded in real court filings, a dataset of 1,300 brief excerpts with injected hallucinations, and evaluate an agentic system integrating open legal database searches. We find that the latest models show promise---within an agentic framework GPT-5 achieves moderate recall (82.8%) and F1 (60.5%)---but struggle with more subtle error categories, leaving significant room to grow. Importantly, though, verification is resource-intensive (GPT-5 averages 16.9 steps per excerpt) and information access barriers limit what even the best agents can achieve. Our taxonomy, dataset, and results provide a foundation for building and auditing legal citation checking tools, as well as identifying policy solutions to raise the bar in what is achievable with agentic tools.