Treadstone Associates
Article · Verification

How to catch AI-hallucinated cases before filing

Every citation in a document you file has to be opened. Not searched for, not recognised — opened, read, and matched against the sentence it is supporting.

Treadstone Associates · Updated 2026

Key takeaways

  • • Three checks, in order: does the authority exist, does it say what the draft says it says, and is it still good law. Most fabrications die at the first; the dangerous ones die at the second.
  • • The Federal Court's Notice on the Use of Artificial Intelligence in Court Proceedings urges reliance on well-recognised and reliable sources, and states that the party signing a document bears responsibility for the accuracy and veracity of its contents.
  • • Alberta's three courts issued a joint notice on large language models, listed as NPP 2025-06 in the Court of King's Bench notices to the profession.
  • • None of this is new law. LSO rule 3.1-2 requires a lawyer to perform legal services to the standard of a competent lawyer, and rule 6.1-1 requires complete professional responsibility for the practice.

The short answer

You catch invented authorities by opening every one of them. That sounds trivial until you notice that the traditional workflow never required it — a citation in a junior's memo came from a database, so the existence of the case was never in question and the check was about whether the proposition was fairly stated. Generative drafting removes that guarantee, and a verification habit built for the old workflow checks the wrong thing.

So the routine has three steps and they run in order: existence, content, currency. Each has a different failure mode and each catches a different class of error.

Step one: does it exist

Take the citation as a string and resolve it in a source that would fail if the case did not exist. That is the operative property, and it is worth being precise about it: a search interface that returns ‘nearest matches’ for a nonsense citation cannot confirm anything. Neither can a summary of the case produced by the same class of tool that produced the citation.

The Federal Court's notice makes the sourcing expectation explicit. Under the heading Caution, it urges caution when using legal references or analysis created or generated by AI in documents submitted to the Court, and states that when referring to jurisprudence, statutes, policies or commentaries it is crucial to use only well-recognised and reliable sources — naming official court websites, commonly referenced commercial publishers, and trusted public services such as CanLII. Alberta's courts said the same thing before the Federal Court did: their joint notice, published as NPP 2025-06, Tri-Court Notice to Profession and Public — Large Language Models, is on the Court of King's Bench notices page dated 6 October 2023.

Step two: does it say what the draft says

This is the step that gets skipped and the step that produces the worse outcome. A real case, correctly cited, standing for a proposition it does not support is harder to detect than an invented one and more embarrassing to be caught on, because you cannot claim you were misled about its existence.

The mechanical version of this check is to require a pinpoint. If the draft asserts a proposition and attaches an authority, the reviewer's job is to find the paragraph. If there is no paragraph, the proposition and the authority have not actually been connected by anyone. The Federal Court's notice describes the corresponding principle as “human in the loop”: to ensure accuracy and trustworthiness it is essential to check documents and material generated by AI, and this kind of verification aligns with the standards generally required within the legal profession.

Step three: is it still good law

Currency is the check a model is worst at, because its training has a horizon and it will describe a reversed decision with the same confidence as a leading one. Nothing about AI changes what you already do here; what changes is that you now have to do it for citations you did not choose.

Who carries this, in the rules

It is worth being clear that no Canadian regulator has created a special AI duty. The existing duties already cover it, which is why they bite. The Law Society of Ontario's Rules of Professional Conduct, Chapter 3 define a competent lawyer in rule 3.1-1 as one who has and applies relevant knowledge, skills and attributes appropriate to the matter, including legal research, analysis and application of the law to the relevant facts — and paragraph (k) of that same definition includes “otherwise adapting to changing professional requirements, standards, techniques, and practices”. Rule 3.1-2 then requires the lawyer to perform any legal services undertaken on a client's behalf to the standard of a competent lawyer.

The supervision rule is the other half. Rule 6.1-1 in Chapter 6 requires a lawyer to assume complete professional responsibility for their practice of law and to directly supervise non-lawyers to whom particular tasks and functions are assigned; the commentary adds that a lawyer is required to review the non-lawyer's work at frequent intervals to ensure its proper and timely completion. A drafting tool is not a non-lawyer, but the standard it sets — complete professional responsibility, work reviewed at frequent intervals — is precisely the standard a firm should apply to output it did not write. British Columbia's Code of Professional Conduct, Chapter 3 and the Law Society of Alberta's professional conduct resources are the corresponding starting points in those provinces.

Worked example (illustrative)

A five-lawyer litigation boutique in Ontario allows generative drafting for first-pass research memos and forbids it for anything filed. In practice the boundary leaks, because a memo becomes a factum by editing.

The firm replaces the prohibition with a citation register. Every authority that appears anywhere in a file is logged once with the name of the person who opened it, the date, and the pinpoint paragraph relied on. Nothing is filed with an authority that has no register entry. The register is not a technology control; it is a bookkeeping control, and that is why it works — it is indifferent to how the draft was produced.

The measurable is the number of authorities in filed documents with no register entry, which should be zero, and the number of register entries where the pinpoint was corrected on verification, which will not be zero and is the interesting number.

If it goes wrong

The consequences sit with the signing lawyer. Treadstone Law's note on when costs can be awarded against a lawyer personally sets out the Ontario position, and its explanation of how expert evidence is admitted at an Ontario civil trial is a useful reminder that the Federal Court notice treats expert reports differently — its declaration requirement does not apply to them, on the Court's understanding that disclosure of AI use belongs in the summary of methodology under the Expert Witnesses Code of Conduct referred to in Rule 52.2 of the Federal Courts Rules.

Questions we get asked

Is there a tool that verifies citations automatically?
Tools can resolve a citation string against a database, which handles step one. Steps two and three require reading, and the whole point of the failure is that a machine's reading is what you are checking.

Does verification have to be done by a lawyer?
It has to be supervised by one. LSO rule 6.1-1 puts complete professional responsibility on the lawyer and requires direct supervision of assigned tasks.

What about client confidentiality when checking?
Verification usually involves no client information at all — you are opening a public authority. The confidentiality question belongs to the drafting step, and the Privacy Commissioner's principles for generative AI are the right starting point for it.

See where AI pays off first in your firm.

A 30-minute call is enough to tell you whether AI pays for itself here.