How to Verify an AI Legal Citation Before You File It

Opposing counsel emails you. They cannot find one of the cases in your brief. They have checked, and they are asking, politely, for a copy.

You go and look. The case has a name, a citation, a court, a year, and a quoted passage that says exactly what you needed it to say. It reads like law. It does not exist.

That is not a hypothetical. It has now happened often enough on both sides of the border to produce reported decisions, and the pattern is the same every time. Nobody set out to mislead the court. Someone ran a search, got an answer that looked right, and filed it.

What the Canadian decision actually held

In Zhang v. Chen (2024), a family lawyer filed an application relying on two cases that ChatGPT had invented. Opposing counsel could not locate them, asked for copies, and the fabrication surfaced.

The costs ruling is the part worth reading twice. Justice Masuhara declined to order special costs, writing: "In my view, the circumstances do not justify the imposition of a special costs award against Ms. Ke", and finding that she had no intention to deceive (Zhang, 2024, para. 32).

Good faith was accepted. Read what it did not prevent. She was ordered to pay the other side's ordinary costs personally, to compensate them for the hours spent chasing authorities that were never there. There is a published judgment with her name on it. The Law Society took an interest.

Why the citation is the part that breaks

A language model produces the most plausible next piece of text. Case citations are the most patterned writing in the profession: a style of cause, a year, a court abbreviation, a paragraph number. That regularity is exactly what a model reproduces well, which is why a fabricated citation looks more convincing than fabricated reasoning does.

The judgment itself notes the same lesson had already been published elsewhere, referring to "the widely publicized case of a lawyer in New York who relied on fake cases from ChatGPT" (Mata, 2023). In that matter the lawyer did check. He went back to the model and asked whether the cases were real, and it told him they were.

That is the step to unlearn. A model asked to confirm its own output will usually confirm it, because agreeing is the most plausible continuation of that conversation. It has no separate memory of what it made up. Asking the tool that produced a citation whether the citation is genuine is not verification. It is the same guess, twice.

The ninety second check

Verification is not difficult. It is a habit that has to survive deadline pressure.

Open the case yourself, in whichever database you use, or on the court's own site. Not a summary of it. The judgment.

Match the style of cause, the year, the court and the citation number. A fabricated cite frequently pairs a real case name with the wrong year, or a real citation number with a different case.

Find the paragraph you are relying on. A real case can still be cited for something it never said, which is the harder error to spot and the one that survives a superficial check.

Check it is still good law. Existing and being current are two different questions.

If you cannot open it, do not file it. No exceptions, however good the quote is.

That is the whole thing. The reason it fails is never that it is hard. It fails at eleven at night before a filing, on the citation that supports your best paragraph.

Where the duty actually sits

Nothing in either decision says a lawyer may not use these tools. Both say the lawyer remains responsible for what is filed. The signature on the document is the assurance, and it always has been. A research tool has never been able to carry that, and the arrival of a more articulate one does not move it.

Treat model output the way you would treat a memo from a bright student who has never been caught being wrong. Worth reading. Never filed unread.

At Kwata Team we build on that assumption. Juris checks every authority against the source before it reaches you, links the judgment so verifying it means opening the case rather than hunting for it, and says so plainly when no authority exists instead of inventing one. The duty stays exactly where it has always been, with you. The tool's job is to make meeting it take seconds rather than an evening.

The habit matters more than the tool. Open the case. Read the paragraph. Then file.

References

Mata v. Avianca, Inc., No. 22-cv-1461 (S.D.N.Y. June 22, 2023). Referred to in Zhang v. Chen (2024) at para. 27.

Zhang v. Chen, 2024 BCSC 285. Supreme Court of British Columbia. Full judgment

Kwata Team