← LegalCite blog · Home

1,667 court decisions and counting: the state of AI-fabricated citations (July 2026)

July 2, 2026 · Every figure in this piece links to its source.

The most complete public record of what happens when AI-invented law reaches a courtroom is the AI Hallucination Cases database, maintained by Damien Charlotin, a research fellow at HEC Paris. It tracks court and tribunal decisions worldwide in which a judge engaged — more than in passing — with AI-generated hallucinations in a filing: citations to cases that do not exist, quotations that appear in no opinion, and real citations attributed to the wrong case.

As of July 2, 2026, the database lists 1,667 decisions. Industry coverage put the count at roughly 1,227 in early 2026 — meaning courts have added on the order of 440 new decisions in about half a year, several every working day.

Two things about that number are easy to miss. First, it undercounts by construction: the database only records instances where a court caught and addressed the problem in a written decision. Fabricated citations that slipped through, were caught by opposing counsel and quietly withdrawn, or never drew a written order don't appear. Second, the consequences attach to names. These decisions identify the filing attorneys, and many impose sanctions, fee awards, or referrals to disciplinary bodies — all of it public, searchable, and permanent.

The failure mode that does the damage

The obviously fake cite — Smith v. Jones, 999 U.S. 9999 — is actually the easy case; any lookup catches it. The dangerous one is the mismatch: a citation that resolves to a real reporter volume and page, attached to the wrong case name or the wrong proposition. It survives a glance because the citation format is perfect and the case at that address exists. It fails only when someone pulls the opinion — and in the sanctions decisions, that someone is usually the judge.

A verification checklist that scales with AI drafting

  1. Existence: does every cited case exist in the public record? (Automatable.)
  2. Match: does the citation resolve to the case name your draft claims? (Automatable — and this is the layer that catches the sneaky failures.)
  3. Quotes: does every quotation actually appear in the opinion? (Pull the opinion; read it.)
  4. Good law: has the case been overruled or limited? (Use a citator — Shepard's/KeyCite. No existence check substitutes for this.)
  5. Proposition: does the case actually support the point cited? (Only a lawyer can do this one.)

Steps 1 and 2 are mechanical, high-volume, and exactly where AI drafting fails most often — which makes them the natural place for automation. Steps 3–5 remain professional judgment, and any tool that claims otherwise is overselling.

LegalCite automates steps 1 and 2: upload a brief and every citation is checked against public court records, with a printable verification report for your file. Free, no signup, your document never leaves your browser — try Brief Check.

LegalCite verifies that citations exist in public court records (CourtListener) and match the case names attributed to them. It does not assess good-law status or holdings, and nothing here is legal advice. Sources: AI Hallucination Cases database (count verified July 2, 2026); PlatinumIDS analysis (early-2026 count).