Disclosure: DROZlegal publishes this guide and builds a practice-automation product for Canadian law firms, including the research module described below. Claims about court decisions and third-party research are drawn from named court citations and peer-reviewed or named-outlet sources, cited throughout; verify current details directly against those sources before relying on them in practice.
A verified-citation problem, not a hypothetical one
Canadian courts are no longer treating AI-fabricated citations as a novelty. In Ko v. Li, 2025 ONSC 2766, Justice Myers confronted a family-law factum containing case citations that couldn't be retrieved on any caselaw service and hyperlinks that led to unrelated decisions or 404 errors — hallmarks of AI hallucination. Counsel was ordered to show cause for why she should not be held in contempt of court; she acknowledged relying on AI-generated content without verifying it, withdrew the factum, and apologized, and a follow-up decision (2025 ONSC 2965) resolved the contempt issue on that basis. A later case conference in the same matter (2025 ONSC 6785) reopened it after her earlier account of who had prepared the factum changed, and the court referred the matter to the Attorney General.
That is one named example in a fast-growing pattern — Canadian courts and tribunals have been addressing AI-hallucinated citations since the first reported case in early 2024, and the consequences have moved well past a judicial warning. In June 2026 the Law Society Tribunal ordered a licensee to pay $31,150 in costs to the Law Society in a proceeding where his uncorrected AI-fabricated citations added to the complexity (Mazaheri v. Law Society of Ontario, 2026 ONLSTH 112) — reported in trade coverage as the largest AI-citation costs award by a Canadian court or tribunal to that point.
A peer-reviewed Stanford RegLab study found that, in May 2024 testing, Lexis+ AI hallucinated in 17% of test queries and Westlaw AI-Assisted Research in 33%, against 43% for an ungrounded general-purpose model (GPT-4). — Magesh et al., "Hallucination-Free? Assessing the Reliability of Leading AI Legal Research Tools," 22 J. Empirical Legal Stud. 216 (2025).
Read that stat carefully: those are leading, commercially sold, retrieval-grounded legal research products — not a hobbyist chatbot — and they still produced wrong or unsupported citations in roughly one query out of every three to six. Two caveats matter. Those figures are a snapshot of specific product versions tested in May 2024, and later third-party benchmarks report different numbers as the products change — which is itself the point: a hallucination rate is a moving property of a particular build, not a permanent guarantee about a vendor. And grounding case-law answers in a real database measurably helps; it does not, by itself, make a tool safe to cite from without checking.
What "grounded in case law" actually needs to mean
Plenty of legal-AI marketing says a tool is "grounded" or "built on real case law." That phrase covers a wide range of actual engineering, and the difference matters to what ends up in a factum. A general-purpose chat model recalls case names from its training data — a statistical guess at what a citation should look like, with no live connection to whether the case exists or says what the model claims. Nothing stops it from inventing a plausible-sounding citation with confidence.
A verified-citation research tool works differently. It retrieves from a maintained, current corpus of real case law and legislation, and it answers only from what that corpus actually contains — the same architectural approach CanLII itself adopted when it launched its own AI-powered search tool, Search+, in February 2026, designed to answer exclusively from CanLII's own case law, legislation, and commentary rather than an open-ended model's memory. That's the difference between a tool that can hallucinate a case and one whose answer is only as good as what's actually in the database it's allowed to search.
What DROZlegal's research module is actually built on
DROZlegal's research capability is built around that same verified-retrieval principle, across three connected pieces:
- A verified-citation case-law corpus, ingested from CanLII, with touch-driven automatic ingestion refreshing a record when it is used rather than leaving the firm on a static, aging snapshot.
- A legislation database covering 19 Ontario and federal statutes and regulations — including the Construction Act, the Courts of Justice Act, the Planning Act, the Land Titles Act, the Canada Business Corporations Act, and the Bankruptcy and Insolvency Act — so a drafting or research question can be checked against the actual statutory text, not a model's recollection of it. (Automatic monitoring for legislative amendments is a planned, not yet shipped, capability — still worth confirming manually against CanLII or e-Laws for anything time-sensitive.)
- Matter-scoped case chat, a research assistant that answers questions against a specific case's own documents together with the case-law and legislation corpus — not the open internet, and not a different matter's files.
None of that makes the output self-certifying. It means a citation DROZlegal's research module surfaces traces to an actual retrievable record in a maintained corpus, which is a meaningfully different starting point than a general chat model's best guess — and still not a substitute for a lawyer opening the case and reading it. Document-level AI carries the same discipline: our guide to AI legal document review software walks through what that layer actually catches in contracts and discovery material, and where a human still has to sign off, the same way a lawyer still has to read the case a research tool surfaces.
Four questions to ask any legal-research AI vendor
"Grounded in case law" is a marketing claim until a vendor answers these in writing. Ask every legal-research AI vendor — DROZlegal included — the same four questions before a citation from their tool reaches a filing:
| Question | Why it matters |
|---|---|
| What database does it actually search? | A named, current corpus (e.g., CanLII) is verifiable; "case law" alone is not. |
| How current is that corpus, and how is it kept current? | A static snapshot from a prior year misses recent decisions and amendments; automatic ingestion is a different guarantee than a manual, occasional update. |
| Can every citation be traced to a real, retrievable record? | If the answer isn't a direct "yes, every time," treat every citation as unverified until you check it yourself. |
| Does it ever answer from outside that corpus? | A tool that falls back to open-ended model recall when its own database comes up empty can hallucinate exactly where you'd trust it most — on the case that doesn't quite exist. |
These four sit alongside the broader vendor questions worth asking about any AI product a firm buys — training use, retention, processing location, and unsupervised authority — covered in full in Is AI Safe for Law Firms? 2026 Vendor Checklist. A research tool that answers the citation-specific questions well but can't answer the general vendor-safety questions isn't actually safe to deploy, whatever its hallucination rate.
The verification duty doesn't go away
Nothing above changes the lawyer's own obligation. A verified-citation architecture reduces how often a tool hands a lawyer a fabricated case; it does not eliminate the need to open that case and confirm it says what the tool claims before it goes in front of a judge. Ko v. Li is, at bottom, a story about that verification step being skipped — not about which AI tool was used. The Canadian Bar Association and Thomson Reuters flagged the trajectory early: in their first joint state-of-the-legal-market reports, released in March 2024, 26% of Canadian law firm lawyers were already experimenting with generative AI, against 6% of government lawyers. Those are 2024 figures and adoption has moved since — which is exactly why the number that matters to a firm is not its adoption rate but whether anyone is still reading the cases line by line.