AI Hallucinated Cases: How Canadian Lawyers Can Use AI Without Getting Burned

Ask a general-purpose AI chatbot a legal question and it will often answer with confidence, complete with case names, neutral citations, and quoted passages. The answer reads exactly like competent legal research. The problem is that some of those cases may not exist. The citations follow the right format, the party names sound plausible, the quotes sound judicial — and none of it is real. This failure mode is called hallucination, and it has become the single biggest professional risk of using AI in legal practice.

This post explains what hallucinated cases are, why the technology produces them, what courts have done about lawyers who filed them, and — most importantly — a practical verification checklist that lets you use AI safely. Because the answer to hallucination is not to avoid AI; it is to use AI that is built to be checkable, and to check it.

What is an AI hallucinated case?

A hallucinated case is a citation, a case name, a holding, or a quotation that an AI model generated but that does not correspond to any real decision. Hallucinations come in degrees, and the subtle ones are the most dangerous:

• Entirely fabricated cases — a style of cause and citation that match no decision in any database.
• Real case, wrong content — the case exists, but the AI attributes a holding or a quote to it that appears nowhere in the judgment. These pass a quick "does it exist" check and fail only when someone reads the actual decision.
• Mangled citations — a real case with the wrong year, court, or citation number, or two real cases blended into one.
• Stale law presented as current — a real case described accurately, but one that has since been overturned or superseded, presented as if it still governs.

What makes all of these treacherous is that they are formatted perfectly. The model has read enormous amounts of legal writing, so its fabrications look exactly like the real thing. Plausibility is the product; accuracy is incidental.

Lawyers have been sanctioned for filing fake cases

Since generative AI chatbots became widely available, courts in several jurisdictions — including Canada — have confronted filings containing AI-invented citations. The pattern is remarkably consistent: a lawyer under time pressure asks a chatbot for supporting authority, receives confident answers with realistic citations, files them without checking, and opposing counsel or the court discovers that the cases do not exist. The consequences have ranged from public embarrassment and cost awards to disciplinary referrals, and judges have made clear that "the AI did it" is not a defence — verifying cited authority is a core professional obligation that predates AI and survives it.

Canadian courts and regulators have responded too. Some Canadian courts now require parties to disclose when generative AI was used to prepare materials, and law societies have issued guidance reminding lawyers that technological competence includes understanding the limits of the tools they use. The direction of travel is clear: AI use is permitted, unverified AI output is not.

Why AI invents cases: prediction versus retrieval

Hallucination is not a bug that better prompting fixes. It follows directly from how large language models work.

A general-purpose LLM is a next-token predictor. Given the text so far, it generates the most statistically plausible continuation, one token at a time, based on patterns learned during training. When you ask it for cases on a legal issue, it is not looking anything up. It is generating text that resembles the answer to your question. Because it has seen thousands of real citations, it knows precisely what a citation should look like — the year, the court abbreviation, the number, the "R. v." — and it can produce a flawless-looking one whether or not any underlying case exists. The model has no internal distinction between remembering and inventing; both are the same act of prediction.

Retrieval is the opposite architecture. A retrieval-augmented system first searches a database of real documents, pulls back actual source text, and then generates its answer from that retrieved text — with the constraint that claims must trace to the sources. The citation exists because the system found the document, not because the citation was statistically likely. This architectural difference, not model size or prompt cleverness, is what separates AI tools that are safe for legal research from AI tools that are not. Our guide to AI legal research in Canada covers this distinction in more depth.

The fix: retrieval from a verified corpus, plus quote verification

A properly built legal AI tool addresses hallucination at two layers:

1. Closed-corpus retrieval. The system answers only from a curated database of real decisions and legislation. If the corpus does not contain relevant authority, the honest output is "no strong authority found" — not an invented case. Every citation in the answer is, by construction, a document that exists.
2. Mechanical quote verification. Even with real sources, a model can paraphrase loosely or misattribute. The second layer checks every quoted passage character-for-character against the source judgment before it reaches you, so a quote in the output is a quote that actually appears in the decision.

Together these convert the AI from an author into a research assistant whose work arrives pre-checked and — crucially — checkable, because every claim links to a real document you can open and read.

A verification checklist for any AI output

Whatever tool you use — a general chatbot, a legal AI product, or a memo from a student — the same four checks apply before any authority goes into your work product. They mirror what careful lawyers did long before AI; AI just makes skipping them more tempting.

1. Does the case exist? Look up the citation in a primary source such as CanLII. Check that the style of cause, year, court, and citation number all match. A near-match — right name, wrong year — is a red flag, not a rounding error.
2. Does it say what is claimed? Open the judgment and find the proposition in the actual text. Verify quotes verbatim and confirm the proposition is part of the court's ratio decidendi rather than obiter, a party's submission, or a dissent.
3. Is it still good law? Note up the case: review later decisions that cite it and confirm it has not been overturned, distinguished into irrelevance, or displaced by legislation. A classic example of why this matters: the framework for trial delay changed fundamentally when R. v. Jordan, 2016 SCC 27 replaced the previous approach — authority that was solid one year was superseded the next.
4. Is the court binding? Confirm the decision comes from a court whose rulings bind (or at least persuade) the court you are before. A trial-level decision from another province is not the same as your own Court of Appeal.

If a tool makes these four checks fast — or performs some of them for you — it is fit for legal work. If checking its output takes longer than doing the research yourself, it is not saving you anything.

How Casescout's verified mode automates the checks

Casescout was built around the retrieval architecture described above. You ask a research question in plain English and get an answer with citations to real Canadian decisions. Under the hood:

• Every answer comes from retrieval, never from model memory. Casescout searches a curated Canadian corpus — Ontario criminal, civil, and family case law plus Canadian legislation, including the Criminal Code — and answers only from what it finds. A case cited in a Casescout answer is a case that exists, which resolves check one by construction.
• Verified research mode reads the full judgments. Rather than working from snippets, verified mode reads the complete decisions and extracts verbatim quotes that are mechanically checked against the source text, addressing check two: the quote is really in the judgment, and you can click through to see it in context.
• You stay in the loop for judgment calls. Currency and bindingness (checks three and four) still deserve a lawyer's eye, and because every citation links to the real decision, noting up from a Casescout answer takes minutes, not hours.

Casescout is free to start, with no credit card required. For how it compares with other options, see our roundup of the best AI legal research tools in Canada.

Frequently asked questions

Can lawyers use ChatGPT for legal research?

Lawyers can use general chatbots for brainstorming, summarizing documents they supply, or drafting prose — tasks where they provide the source material. Using one as a source of legal authority is where lawyers get burned, because a general LLM generates plausible citations rather than retrieving real ones. Any authority it names must be independently verified before it goes anywhere near a filing, and many lawyers find that verification burden erases the time savings.

How can I tell if a case citation is fake?

Look it up in a primary source like CanLII. If the citation resolves, confirm the style of cause, year, and court all match what the AI told you, then read the judgment to confirm it actually supports the stated proposition. Fabricated citations often fail at the first step; the subtler failures — real case, invented holding — only surface when you read the decision itself.

Do Canadian courts allow AI-prepared court filings?

Courts have not banned AI, but they hold lawyers fully responsible for what they file, and some Canadian courts require disclosure when generative AI was used to prepare materials. Check the practice directions of the specific court you are before, and treat verification of every cited authority as non-negotiable regardless of disclosure rules.

Does using a legal AI tool eliminate hallucination risk?

A retrieval-based tool with a verified corpus eliminates the worst failure — citing cases that do not exist — and mechanical quote checking eliminates invented quotations. Professional judgment about currency, bindingness, and application to your facts remains yours. The realistic goal is not zero human review; it is review that takes minutes because every claim is linked to a real, readable source.

What should I do if I find a suspected fake citation in materials?

Verify it against a primary source first — some "fake" citations are merely mangled versions of real ones. If it is genuinely unverifiable, raise it promptly: with the drafter if the materials are your side's, or with the court and opposing counsel as candour obligations require if it has already been filed. For a grounding in research fundamentals, see our guide on how to do legal research in Canada.