A language model doesn't look anything up. When you ask it for supporting authority, it predicts text that looks like a citation — a plausible case name, a real-format reporter number — without any guarantee the case exists.
An LLM generating Smith v. Jones, 999 U.S. 9999 (2050) is doing exactly what it was trained to do: produce statistically likely tokens. "United States Reports" citations follow a rigid volume U.S. page shape, so the model fills the slots convincingly. The result is a citation that passes a human eye but points to nothing. Worse, it often pairs a real reporter slot with the wrong case name — a "mismatch" that's even harder to spot.
Lawyers who filed AI-drafted briefs without checking the authority have been sanctioned when opposing counsel and judges discovered the cited cases didn't exist. The cost isn't just embarrassment: it's professional discipline, lost cases, and malpractice exposure. And it scales badly — a single brief can carry dozens of citations, and checking each by hand against a database is exactly the kind of work people use AI to avoid.
The fix isn't "trust the model less." It's to verify every citation against real court records, automatically, in the drafting loop.
Real U.S. opinions are public. CourtListener (run by the non-profit Free Law Project) holds millions of them, addressable by citation. LegalCite reads your text, detects every citation by its reporter pattern, and resolves each against those records:
There is no fabrication path: if a citation can't be resolved, LegalCite says so — it never invents a match.
LegalCite is a REST API and an MCP server. Agents call verify_citations(text) on any draft before relying on it; drafting tools call /api/verify on every paragraph. Free tier is 20 checks/day, no signup.
LegalCite verifies whether a citation exists in public court records and returns its metadata. It does not assess legal correctness, holdings, or good-law status, and is not legal advice.