Catch AI-hallucinated citations
before your brief reaches a judge.
Generative AI fabricates citations that pass the eye test — proper reporter, plausible parties, plausible year — and don't exist. Or they exist, but the opinion says nothing about your proposition. CiteVerify reads the actual opinion behind every cite and flags both failure modes in one report, in under a minute.
Check a brief now →What other checkers miss
Five phases. One report. Under a minute.
Every hallucination mode a court has sanctioned an attorney for, checked in one pass.
Does the case exist?
Every citation is matched against CourtListener's 8.3M-decision corpus and the Free Law Project citation index. Fabricated cases — the most-sanctioned failure mode — are flagged before you file.
Does the opinion actually say that?
We pull the full opinion text and check whether the cited authority stands for the sentence it is cited for. This is what KeyCite and Shepard's don't do. This is where most sanctions come from.
Is the citation formatted right?
Bluebook 21st and 22nd edition rules, jurisdictional variants, pin cites, short forms, Id., and signal placement. Format errors caught with copy-to-clipboard fixes.
Is this a leftover model artifact?
Placeholder tokens, orphaned brackets, "[insert citation]", generic party names, phantom reporter volumes — the drafting-model tells that show a human never checked the passage.
Is the authority still good law?
Good-law check across CourtListener's treatment graph. Overruled, abrogated, distinguished, questioned, or narrowed — surfaced with the citing decisions.
Does the statute say that?
Statutory citations verified against the current codified text, with amendment-date awareness so a 2019 citation is checked against the 2019 version, not today's.
What other checkers can't catch
Case-existence checkers verify half the failure modes. Treatment tools verify a different half. Only proposition support checks whether the citation says what you claim it says.
| CiteVerify | KeyCite / Shepard's | Standalone cite checkers | Generic LLMs | |
|---|---|---|---|---|
| Case exists in a real corpus | ✓ | ✓ | ✓ | — |
| Opinion text supports the proposition | ✓ | — | — | — |
| Bluebook format check | ✓ | partial | ✓ | — |
| Treatment / good-law | ✓ | ✓ | — | — |
| AI-drafting artifact detection | ✓ | — | — | — |
| Statute verification with amendment awareness | ✓ | ✓ | — | — |
| Per-document pricing from $10 | ✓ | — | varies | n/a |
Real sanctions. Real hallucinations.
These are all documented court decisions where AI-fabricated citations reached the bench. Every pattern below is exactly what CiteVerify's five phases are built to catch.
Fabricated opinions passed as controlling authority
Attorneys sanctioned for filing a brief citing six cases that did not exist. Bluebook-perfect format, plausible party names, no case behind them. Caught by Phase 1.
Court disclosure of AI use required
Circuit court confirmed judges may require disclosure of AI use and an explanation of how citations were checked. Documented verification is now a defense.
Real cases cited for propositions they never held
The fastest-growing sanction pattern. The case is real. The opinion says nothing about the proposition. Cite-existence checkers pass it. Caught by Phase 2.
Frequently asked
What counts as an AI hallucination in a legal citation?
Two patterns. First: the case doesn't exist — the model invented the reporter, the year, or the parties. Second: the case exists, but the cited opinion says nothing about the proposition you cited it for. Both are grounds for sanctions. Both are caught by CiteVerify's Phase 1 and Phase 2 respectively.
How is this different from Westlaw KeyCite or Lexis Shepard's?
KeyCite and Shepard's tell you whether a case is still good law. They do not open the opinion and check whether the sentence you cited it for is actually supported by the opinion text. CiteVerify does both — treatment status in Phase 4, and proposition support in Phase 2, which is where most modern sanctions originate.
Do you send my brief to a third-party AI?
Verification runs on Amazon Bedrock in a private VPC via the ClientShield platform. No model training on your documents. Full details on the ClientShield security page.
What corpus do you check against?
CourtListener and the Free Law Project — 8.3 million decisions across 3,355 jurisdictions. The same open dataset used by federal courts' citation initiatives, Harvard Law Library, and public-access legal research projects. No proprietary lock-in.
How much does it cost?
Per-document from $10 (1–10 pages) up to $30 (36–50 pages). Cite-Check Pro is unlimited briefs at $99/month. Track C of the ClientShield platform includes CiteVerify unlimited plus anonymization, private-AI research, and Discovery Analyst. See pricing.
One bad citation can end a career.
Every citation extracted. Every opinion actually read. Every hallucination flagged with the location, the reason, and the fix. Before the brief reaches the court.
Check a brief now →