Can AI be used for legal research in Canada?

Yes, for locating and summarising real documents. Open corpora such as A2AJ publish bulk Canadian court decisions that AI assistants can query directly. AI is not reliable for recalling case law from memory: general-purpose assistants generate citations that appear correct but do not exist, and courts have sanctioned lawyers who filed them. Verify every citation against CanLII or the court's own site before relying on it.

A2AJ is the Access to Justice AI project, which publishes bulk, openly licensed Canadian legal data and exposes it to AI assistants. As of August 2026 its holdings include the Supreme Court of Canada from 1877, the Ontario Court of Appeal from 1998, the Federal Court and Federal Court of Appeal from 2001, and the Canadian Human Rights Tribunal from 2003, among roughly thirty datasets.

Does A2AJ include HRTO decisions?

No. The corpus does not hold Human Rights Tribunal of Ontario decisions, and it holds no Ontario Superior Court or Divisional Court material. Because the Divisional Court hears every judicial review of an HRTO decision, that reviewing layer sits outside the open record. HRTO and Divisional Court decisions must be found on CanLII or the Tribunal's own publication.

How do I check whether an AI-generated case citation is real?

Search the neutral citation itself, such as 2013 SCC 19, on CanLII or the court's own website rather than searching the case name. Confirm the parties and year match, open the specific paragraph you intend to rely on and read it, and check the citing-cases list to confirm the decision was not overturned. If a citation fails any step, discard it entirely.

AI Legal Research Tools Canada: What Works, What Invents Citations

Can AI help with a legal problem in Ontario?

For finding and reading material, yes — there is now a genuinely open corpus of Canadian court decisions that both people and AI tools can query directly. For telling you what the law is, no: general-purpose chatbots invent citations that look correct, and lawyers have been sanctioned for filing them. The workable method is to use AI to locate and summarise real documents, then verify every citation against the source before relying on it.

This page sets out what each tool actually covers, what it cannot see, and a verification method that takes about two minutes per citation.

The open corpus: A2AJ

The Access to Justice AI project publishes bulk, openly licensed Canadian legal data, and exposes it through an interface that AI assistants can query directly. It is the single most useful development for self-represented people and small practices in years, and almost nobody knows it exists.

We queried its coverage on 7 August 2026. These are its actual holdings — not a description of them.

Dataset	Court or tribunal	Earliest	Latest	Documents

Fourteen of roughly thirty datasets shown, selected for relevance to Ontario human rights and police accountability work. Retrieved 7 August 2026; holdings grow continuously.

What it does not hold, and why that matters here. The corpus carries the Ontario Court of Appeal from 1998 and the Supreme Court from 1877 — but no Ontario Superior Court and no Divisional Court. Since the Divisional Court is where every judicial review of an HRTO decision goes, the entire reviewing layer of Ontario human rights law sits outside the open record. It also holds no HRTO decisions at all: for those you still need CanLII or the Tribunal's own publication.

That absence is not a criticism of the project. It is the same shape of problem this site documents — the record that exists is not the record you can search.

What each tool is actually for

A2AJ

Best for: full decision text, bulk analysis, appellate research, and letting an AI assistant read real judgments rather than recall them.
Cannot: HRTO, Divisional Court, Ontario Superior Court.
Cost: free, open licence.

CanLII

Best for: HRTO decisions, Divisional Court, everything A2AJ lacks. Citator shows what cited what.
Cannot: be scraped — its terms and robots.txt prohibit automated collection, and the API returns metadata only.
Cost: free to read.

Claude, ChatGPT, Gemini

Best for: explaining a document you paste in, drafting a chronology, turning notes into a timeline, plain-language summaries.
Cannot: be trusted for citations, holdings or deadlines from memory.
Cost: free tiers exist.

CLEO Steps to Justice

Best for: verified plain-language Ontario procedure, in many languages, with Guided Pathways that help complete forms and draft letters.
Cannot: analyse your specific facts.
Cost: free.

Tribunal & court sites

Best for: the authoritative version of a rule, form, practice direction or deadline. Always the last word.
Cannot: be searched well.
Cost: free.

Paid research platforms

Best for: comprehensive coverage and professional citators.
Cannot: be afforded by most self-represented people.
Cost: substantial subscription.

The platforms, and who each is realistically for

The landscape changed fast in 2026. Two things now matter more than the rest: CanLII added AI search, and an open corpus arrived that works inside general assistants. Both are free.

Free, and usable by anyone

Tool	What it is	Limits
CanLII Search+	AI search layered on CanLII, rolled out in the first quarter of 2026. The best free database of Canadian decisions, now with AI on top. Requires a free My CanLII account.	Daily use limits. Some lawyers find these a hindrance for daily work.
A2AJ	Open bulk Canadian legal data that works inside Claude or ChatGPT, so an assistant reads real judgments rather than recalling them.	No HRTO, no Ontario Superior Court, no Divisional Court. See the coverage table above.
Steps to Justice	CLEO's verified plain-language Ontario procedure, in many languages, with Guided Pathways that help complete forms and draft letters.	Cannot analyse your facts.
General assistants Claude, ChatGPT, Gemini	Excellent at transforming documents you supply — chronologies, summaries, plain-language translation, consistency checks.	Unreliable for recalling case law. See the failure mode below.

Paid, and aimed at the profession

Worth understanding even if you will never buy one — because the other side may be using them, and because the gap between what they do and what a free tool does tells you where to spend your own effort.

Lexis+ with Protégé

What it does: research, drafting and analysis in one workflow, grounded in Lexis content with linked citations rather than generated from memory. Includes summarise-a-case and upload-a-document-and-ask. LexisNexis renamed Lexis+ AI to Lexis+ with Protégé in February 2026; existing users carried over.

What that means in practice: ask "what is the test for family status discrimination in Ontario and how has it been applied since 2016" and get a memo with citations you can click through to the full text. Free for academic subscribers — if you are a student, or can reach a university law library, this is the single biggest capability gap you can close for nothing.

Westlaw Edge Canada with CoCounsel

What it does: the Thomson Reuters equivalent. CoCounsel was formerly the standalone Casetext product and is now folded into the Westlaw ecosystem.

In practice: document review across a large set of files, deposition and hearing preparation, timeline extraction from a document bundle, and a citator that tells you whether an authority still stands. The citator is the part with no free equivalent of comparable depth.

Alexi

What it does: purpose-built for litigators, with a strong Canadian focus. Its signature output is a structured research memo on demand — pose a complex litigation question, receive a memo with the answer and supporting authority.

In practice: a different job from case-law search. You are not browsing; you are commissioning a first draft of the research a junior would have produced. Aimed at the memo-and-motion workflow.

vLex Vincent

What it does: grounding across a very large multi-jurisdiction database — vLex spans more than 110 jurisdictions. vLex became part of Clio following an acquisition completed in late 2025.

In practice: the strongest option when the question crosses borders, or when you want serious grounding without committing to the Westlaw or Lexis duopoly.

Clio

What it does: practice management — matters, billing, documents, client intake — with AI layered on. Clio acquired the Canadian legal data company Jurisage in June 2026; Jurisage's Compass held a case-law database of more than 470,000 cases.

In practice: not a research tool first. It is the system a small firm runs on, and the AI features sit where the work already is. Relevant here because Canadian legal data now sits underneath it.

Spellbook

What it does: contract drafting and review inside a word processor — clause suggestions, risk flagging, redlining. Named the Canadian Bar Association's exclusive AI contract drafting and review partner in March 2026, in a two-year deal covering roughly 40,000 members with member data stored in Canada and not used to train models.

In practice: transactional rather than litigation. Irrelevant to a human rights application; highly relevant if a settlement agreement is put in front of you. The data-residency term is the part worth noting — it is the question to ask of any tool you put client information into.

What none of them do. Every tool above still requires you to know what question to ask, and none of them will tell you that your limitation period expired last month. The expensive tools compress research time. They do not supply judgement, and they do not carry responsibility for the answer — you do.

For a self-represented person, the honest shortlist is two items. CanLII Search+ to find decisions, and a general assistant to read the ones you have downloaded. Everything in the paid table exists to save billable hours, which is not the problem you have.

Authoritative guidance exists. The Action Committee on Court Operations, through the Office of the Commissioner for Federal Judicial Affairs, has published an overview of the AI features built into CanLII, Lexis+ AI and Westlaw Edge Canada — written for court officials, legal professionals and self-represented litigants. It is a better starting point than any vendor's marketing.

The failure mode, stated plainly

AI assistants generate citations that do not exist. They are fluent, correctly formatted, plausibly named, and entirely fictional. Courts in Canada and elsewhere have sanctioned lawyers who filed them. A fabricated citation is worse than no citation: it destroys your credibility on everything else you said, including the parts that were true.

The failure is systematic rather than occasional, and it gets worse the more specific you are. Asking for "a case about racial profiling at the HRTO" produces plausible-looking case names; asking an AI to read a decision you supply produces reliable summaries. The difference is whether the document is in front of it.

A two-minute verification method

One

Take the citation and search it directly on CanLII or the court's own site. Not the case name — the neutral citation, e.g. 2013 SCC 19.

Two

Confirm the parties and the year match what you were told. A real citation attached to the wrong case is a common variant.

Three

Open the paragraph you intend to rely on and read it yourself. Summaries drift; the paragraph number is often invented even where the case is real.

Four

Check it was not overturned. Use CanLII's citing-cases list. A correct quotation from a reversed decision is still a losing argument.

If a citation fails any of these steps, discard it entirely rather than trying to repair it — a tool that invented one will usually have invented the reasoning around it too.

What this actually looks like in use

Three worked examples from our own research on this site, including where the method failed.

Example 1 · Finding an appellate decision that turned out to matter

Question: does filing a police complaint prevent you from suing later?

Method: queried the open corpus across the Supreme Court dataset for judicial review of police disciplinary findings. Retrieved Penner v. Niagara (Regional Police Services Board), 2013 SCC 19, and read the full text.

Result: a 4–3 decision holding it was unfair to bar the civil claim, with reasoning that does not appear in any plain-language Ontario resource we could find. What it holds, and its three qualifications.

Example 2 · Correcting a comparison we had made ourselves

Question: why did one racial profiling finding produce $10,000 and another $80,000?

Method: pulled the full Court of Appeal judgment in Briggs, 2022 ONCA 823, rather than relying on summaries.

Result: the full text showed the settlement terms, the eighteen months on reserve, and that $30,000 of the larger award sat in heads the Tribunal cannot order at all — which changed the comparison from a headline ratio into something defensible. Reading the source corrected our own analysis.

Example 3 · Where it failed

Question: what did Ontario Boards of Inquiry award for discrimination in the mid-1980s?

Method: corpus queries, then general search.

Result: nothing usable. Pre-1990 Board of Inquiry decisions are not systematically in CanLII and are absent from the open corpus entirely. We could not confirm a figure and said so on the page rather than publishing an unsourced number. When the record does not exist, the correct output is an admission, not an estimate.

Prompts that tend to work better

Weaker	Usually stronger
"Find me cases about racial profiling by police at the HRTO."	"Search the corpus for Ontario Court of Appeal decisions mentioning racial profiling and police, and give me the citations you actually retrieved."
"What's the deadline to file a human rights complaint?"	"Quote section 34 of the Ontario Human Rights Code and link the source." Then check the Tribunal's own site.
"Summarise Briggs v Durham."	"Here is the full text of 2022 ONCA 823 [paste]. Summarise the procedural history and quote the paragraph on the settlement terms."
"Do I have a good case?"	"Turn these notes into a dated chronology with a column for what document would prove each entry."
"Write my application for me."	"Read my draft and list every factual assertion that has no document behind it."

The pattern, so far. These tools are more dependable when the material is in front of them and the task is transformation — summarising, structuring, translating, checking consistency — and less dependable when the task is recall. Giving them documents rather than questions about the law is a good default, not a rule; the tools change quickly and there is plenty of room to find better approaches than these.

What AI will not do for you

Meet your deadline. No tool files anything. The one-year limit under s. 34(1) runs regardless.
Tell you whether you have a case. That requires facts, judgement and someone accountable for the answer.
Preserve evidence. Footage is overwritten in weeks. Send the preservation request yourself, today.
Replace a person. Free advice exists with no income test. Who covers what.
Keep your information private, necessarily. Read the terms before pasting medical records, a child's details, or anything under a publication ban.

← The Record Get your records →