The failure has a signature. A brief arrives citing four authorities. The names sound right. The reporter abbreviations are correct. The volume and page numbers are plausible for the era. The pin cites are precise. And two of the cases do not exist.
This is the most-discussed risk of using AI in legal work, and it is worth being precise about why it happens, what it does and does not tell you, and what a system has to do structurally — not aspirationally — to stop it.
An honest note about the famous example
The case everyone reaches for is Mata v. Avianca, in which a New York federal court dealt with a brief containing citations produced by a chatbot.
Here is what I can confirm from a primary source, and what I cannot. The docket is real and checkable: Mata v. Avianca, Inc., No. 1:22-cv-01461, United States District Court for the Southern District of New York, filed 22 February 2022 — confirmed through CourtListener’s RECAP archive of federal court records. The sanctions opinion itself I could not retrieve from any source I can verify against: CourtListener’s search API returns no matching opinion for the case name, the Caselaw Access Project’s scanned reporters do not extend to that volume, and its citation-lookup endpoint requires credentials I do not hold.
So I am not going to quote it, and I am not going to state what it held as though I had read it. That is a slightly awkward way to open an article about citation integrity. It is also the entire point: the discipline is not “cite carefully”, it is state exactly what you have verified and how far it goes.
What I can offer instead is a judicial observation from a document I do hold. In United States v. Heppner, 25 Cr. 503 (JSR) (S.D.N.Y., 17 February 2026), Rakoff J noted in a footnote that “most published decisions involving generative artificial intelligence have had to do with attorneys’ misuse of that technology”. The pattern is real enough that a federal judge treated it as the background against which a different AI question had to be decided.
Why language models invent cases
A general-purpose language model predicts plausible continuations of text. That is not a bug to be patched; it is the mechanism.
Consider what that means for a citation. The shape of a legal citation is enormously regular — party names, volume, reporter, page, court, year — and a model has seen a great many of them. Producing something with the right shape is easy. Producing the particular one that exists, and that says what the sentence needs it to say, requires the model to have retained a specific fact and to prefer it over a fluent guess.
When it has not retained that fact, the plausible-sounding citation is often more statistically likely than the messy truth, which might be “there is no case directly on this point”. The model is not lying. It is doing exactly what it was built to do, in a domain where “plausible” and “true” must never be allowed to blur.
Note the consequence: fluency is not a signal of reliability here — it is the source of the problem. A hesitant, badly formatted citation would be easier to catch. The dangerous output is the one that looks like it came from a textbook.
Why it matters more in law than almost anywhere else
In most fields a wrong reference is an embarrassment that costs you credibility. In law, authority is the argument.
A contract analysis that says “this clause is likely unenforceable” is worth nothing unless the statute or case behind that statement exists and says what it is claimed to say. Strip out the authority and you are left with an assertion of the same standing as any confident stranger’s opinion. Worse, a fabricated citation is not merely absent support — it is affirmatively misleading, because it invites the reader to stop checking.
And the downstream cost is asymmetric. The person best placed to catch the error is the one least likely to look: the reader who does not know the area, which is precisely the reader who needed the analysis.
The engineering answer: a curated authority bank
The response has to be structural, because you cannot instruct a probabilistic system into being reliable about facts it does not hold. Ours has four parts.
A curated bank of authorities. The statutes and leading cases for each jurisdiction we cover, checked against primary sources — legislation.gov.uk, the National Archives’ Find Case Law, official legislature sites, the Supreme Court of Canada’s own decisions service, and the Caselaw Access Project’s scans of the official US reporters. As at today the bank holds 322 distinct authorities: 73 cases and 249 statutes.
Verbatim citation only. Where the engine cites from the bank, it cites word for word, and the citation is marked as coming from the bank. It is instructed, in the strongest terms available, never to compose a citation.
Visible flags for everything else. If an analysis draws on an authority outside the bank, it is marked for verification and presented as a name and a year — never a fabricated law report reference, and never presented as checked. That flag is the honest signal that the reader is holding a lead, not a conclusion.
An automated check that can fail the build. The US case citations that the archive covers are machine-checked against the official reporters, and the script exits with an error on any mismatch. On the current pass that is seven verified and one out of coverage — a 2015 Supreme Court decision whose bound volume the archive has not reached, which is recorded as “no coverage” rather than quietly treated as either confirmed or wrong.
The limits, stated plainly
Three things this does not do, and pretending otherwise would undercut everything above.
It is a control on citations, not on reasoning. An engine can cite a real statute, correctly, and still draw the wrong conclusion from it. Catching invented authorities is the floor, not the ceiling.
It does not extend to every jurisdiction equally. Some official sources are simply not reachable by automated checking — several national legal databases block scripted access outright — so authorities in those jurisdictions rest on the curator’s own reading rather than an automated check, and that is a different kind of assurance, honestly a weaker one for scale.
It is not a substitute for a practitioner checking the work. The value of the flag is that it tells you where to look. Somebody still has to look.
A different risk, worth knowing about
Fabricated citations are the risk everybody discusses. Heppner raises a second one that gets far less attention, and it points the opposite way — it is about the citations being real and the exchange being discoverable.
The defendant there had, on his own initiative, used an AI platform to work through the case against him, and later gave the material to his lawyers. The court held the documents were protected by neither attorney-client privilege nor the work product doctrine, and the first reason was short: “Because Claude is not an attorney … that alone disposes of Heppner’s claim of privilege.” The second was that the exchanges were not confidential, the court reading the platform’s published privacy policy to reach that conclusion.
Two qualifications matter before anyone over-reads that. It concerned a consumer-facing policy rather than commercial terms, which differ; and the court expressly left open, at page 7, that had “counsel directed Heppner to use Claude”, the platform “might arguably be said to have functioned in a manner akin to a highly trained professional who may act as a lawyer’s agent within the protection of the attorney-client privilege”. Direction by a lawyer may change the analysis. Running the tool yourself and forwarding the output afterwards, on this reasoning, does not.
How to check any AI legal output, including ours
The useful habit is not scepticism in general; it is a specific sequence, and it takes about two minutes per authority.
Ask first whether the citation resolves — does the case or section exist, at that reference, in an official source? For UK material, legislation.gov.uk and the National Archives’ Find Case Law are free and authoritative. Then ask the harder question, which almost nobody asks: does it say what the analysis claims? A real citation attached to a proposition it does not support is more common than a wholly invented one, and it survives the first check.
Then check the date. Statutes are amended and repealed, and a provision that existed is not the same as a provision in force on the day your facts arose.
If a tool cannot tell you which of its citations were verified and which were not, that is the answer to your question about the tool.
None of this makes an AI report legal advice. It makes it something more modest and more useful: a preliminary analysis you can actually check, in which every citation is either verified or visibly flagged. That checkability is the whole difference between decision support and a confident guess. If you are weighing where AI belongs in your own work, the same test applies to briefing a case.
Check your document before you sign it. Legal Case Analyst reads contracts, leases, judgments and legislation under the law of your specific jurisdiction and returns severity-ranked issues with verifiable citations — on screen in minutes, with Word and PDF to download. £12 for the first 100 reports, standard price £24.
This article is general information, not legal advice. No solicitor/attorney–client relationship is created by reading it. For advice on your situation, consult a licensed practitioner in your jurisdiction.
