Since 2023, courts have sanctioned lawyers for briefs built on cases that
never existed. Vera reads any legal text and checks every citation against
the real court record — so a fabricated case, a wrong name, or an overruled
precedent can't slip through.
Ask a legal question. We send it to an AI, then verify every case it cites —
live. Watch it invent authority, and watch us catch it.
The AI's answer — every citation below was
checked live against the court record:
2,046
court cases, and counting, where AI-fabricated
citations were caught in real legal filings
This isn't hypothetical. Since a New York lawyer filed a brief full of
cases ChatGPT invented, courts on five continents have caught the same
mistake over and over — often only because a judge or opposing counsel
checked the citations by hand. Vera does that check in seconds.
1,397United States
218Canada
112Australia
1,693fabricated cases
The case that started it
Mata v. Avianca, Inc.$5,000 sanction
S.D.N.Y. · June 2023 · Judge P. Kevin Castel
Two attorneys submitted a brief citing at least six cases that did not
exist — Varghese v. China Southern Airlines,
Shaboon v. Egyptair, and others. ChatGPT had invented them,
complete with fake quotes and fake internal citations, and even
"confirmed" they were real when asked. The court sanctioned both
lawyers and their firm.
It keeps happening
Park v. Kim2d Cir., 2024 — attorney cited a nonexistent case generated by ChatGPT; referred for discipline
Michael Cohen matterS.D.N.Y., 2023 — fake cases from Google Bard passed to a lawyer, then filed in federal court
Beus Gilbert PLLC v. BYUD. Utah, 2026 — fabricated case law via multiple AI tools; $3,000 sanction plus mandatory AI-ethics training
Aggregate figures from the public
AI Hallucinations legal database
maintained by Damien Charlotin (snapshot: Sept 2026). Individual cases
are drawn from public court records — verify any of them the same way
Vera does: against the source.
We check the law. We never write it.
Vera exists because generative AI has one specific, dangerous failure: it
invents legal authority. It will produce a case name, a reporter citation,
a year, and a confident quotation — for a case that does not exist. Lawyers
have filed these fabrications in real courts, and real judges have
sanctioned them for it.
Every other AI tool answers the way it broke: by generating more text.
Vera does the opposite. It generates nothing. It reads the
citations already in front of it and checks each one against the public
court record — ground truth, not a guess. The result is reproducible and
auditable: run it twice and get the same answer, and follow every link back
to the source opinion.
How it works
1
Extract
Citations are parsed with eyecite, the Free Law Project's deterministic parser. No model decides what counts as a citation.
2
Verify
Each citation is looked up in CourtListener's database of millions of real opinions. We confirm the case exists, that its name matches what was written, and that any quoted language actually appears in the opinion.
3
Check it's still good law
A real case can still be bad law. Vera flags landmark precedents that were later overruled — citing an overturned case today is its own kind of error.
Why you can trust a red flag
Vera is built precision-first: it would rather say “verify this manually”
than falsely accuse a real case. A citation is only called fabricated when
there is a full case name and no case by that name exists anywhere
in the record. Everything uncertain is surfaced as uncertain — never
dressed up as a verdict.
Who's behind Vera
Vera was built by Atharv Pratap Singh Chauhan, a student
at IIT Bombay, out of one conviction: software that people rely on in court
should never answer “trust me.” It should show its work against the record —
every case, every quote, every time.
Vera was created at LexHack 2026. Case data comes from the
CourtListener
API by the nonprofit Free Law Project; citations are parsed with eyecite.
Real-world figures on the “Why it matters” page are drawn from the public
AI-hallucinations legal database maintained by Damien Charlotin.