Legal Marketing

AI Hallucinations in Legal Work: How to Safeguard Your Firm

15+ years in local marketing; Google Ads certified; Shopify Partner.

You've probably seen the headlines. A lawyer in New York filed a brief with six fabricated case citations. A BC lawyer got sanctioned in early 2024 for submitting fake cases hallucinated by ChatGPT in a family law matter. The judge wasn't amused. Neither was the Law Society.

So let's talk about ai hallucinations legal work, because this is one of those topics where the risk is real, the rules are still being written, and the "just don't use AI" advice is already out of date. Your associates are using it. Your paralegals are using it. The question is whether your firm has a system to catch the hallucinations before they hit a courtroom or a client file.

Here's the thing. This article isn't about whether you should use AI. That ship sailed. This is about the operational safeguards a Canadian firm needs so an AI hallucination doesn't cost you $15K, a formal complaint, or your practice. If you want the broader AI strategy piece, check our full guide to legal AI tools for Canadian law firms, and the wider SEO playbook for Canadian law firms covers how AI fits into how clients actually find you.

Quick math on why this matters before we go further. A first-year associate billed at $250/hr who spends two extra hours hunting down a fabricated citation, plus 30 minutes for a partner to review the close call, costs the firm roughly $625 in unbilled time per incident. If that happens twice a month across a six-lawyer firm, that's $15,000 a year leaking out of utilization. That's the loss when nothing actually gets filed. The math gets uglier fast if a hallucinated cite lands in a brief. If you're trying to figure out if ChatGPT is even worth the risk for your practice, we broke that down in ChatGPT for lawyers: 8 use cases and 4 risks.

This one's tactical. What hallucinations actually are, where they show up in legal work, and the exact process to catch them.

What an AI Hallucination Actually Is (In Plain English)

A hallucination is when a large language model (ChatGPT, Claude, Gemini, Copilot, Lexis+ AI, whatever) generates text that sounds confident and correct but is factually wrong. Made up. Fabricated.

The model isn't lying on purpose. It's not lying at all. It's predicting the next most likely word based on patterns from its training data. When it doesn't know something, it doesn't say "I don't know." It guesses. And it guesses in the same confident tone it uses for things it actually knows.

For lawyers, this matters because legal work rewards precision. A case citation is either real or it isn't. A statute section is either 4.2-1 or it's not. An AI will happily generate a citation that looks perfect, in proper McGill Guide format, with a plausible judge name, a plausible year, and a plausible court, and the case does not exist.

The three flavours of hallucination that hurt Canadian firms most:

Fabricated case law. Citations to decisions that don't exist. Or real case names with the wrong ratio. Or real cases that say the opposite of what the AI claims.
Misstated statutes and rules. "Section 7 of the Divorce Act says..." and it doesn't.
Made-up facts about clients or files. When you feed an AI a summary and ask it to draft something, it will sometimes invent details that aren't in your source material. A date. A dollar figure. A party name.

All three are career-ending if they make it into a filing.

Where Hallucinations Show Up in a Real Law Practice

This is where I think a lot of firms underestimate the risk. Most partners picture the dramatic courtroom scenario, the fake Air Canada case, the judicial dressing-down. That's the obvious one. But the hallucinations that actually erode a firm are the quiet ones.

In my experience, across the Canadian firms we've worked with on digital and content, hallucinations tend to show up in five places:

Website content and blog posts. A junior or a marketing coordinator asks ChatGPT to "write a blog post on spousal support in Ontario." The output references a made-up 2023 Ontario Court of Appeal decision. It goes live. A potential client (or worse, another lawyer) sees it. This is the one I see most often, and it's the one the Law Society will notice first because it's public.

Client intake summaries. An AI chatbot summarizes an intake call. It "remembers" details that weren't said. The summary lands in your matter management system. Six months later an associate relies on it. For more on this specific risk, we've got a dedicated breakdown of AI client intake chatbots and UPL risk.

First-draft memos and research. Associate asks Copilot to summarize five cases on a specific issue. Four are real, one is invented. If the associate doesn't pull every single cite, the memo carries a ghost.

Demand letters and client emails. A lawyer dictates the bones, an AI cleans it up, and the AI adds a "for clarity" line that states a legal position that's wrong.

Marketing copy with legal claims. "Our firm has recovered over $50M for clients." Did it? Prove it. Under Ontario Rule 4.2-1(a), every marketing claim must be "demonstrably true, accurate and verifiable." An AI doesn't know what your firm has recovered. It guesses.

That last one connects to an entire compliance minefield. If you're producing any AI-assisted marketing content, read our piece on AI-generated legal content and Law Society rules before you publish another word.

The Canadian Regulatory Reality (It's Moving Fast)

Here's where I'll be honest. The rules are in flux, and anyone who tells you they have a clean, settled answer is bluffing.

What we know as of 2026:

The Federation of Law Societies of Canada Model Code (Rule 3.1 on competence, Rule 5.1 on candour to tribunal) already covers this. You're competent or you're not. You submitted a real citation or you didn't. AI doesn't get you a pass.
The Law Society of Ontario issued guidance in late 2023 and updated it through 2024-2025 reminding lawyers that supervising technology is part of the competence duty under Rule 3.1-2. Translation: if your AI hallucinates and you didn't catch it, that's on you.
The Law Society of BC has been more forward on practice directions around generative AI, particularly after the 2024 Zhang v. Chen sanction. BC firms should read the current Benchers' guidance directly.
Alberta and Saskatchewan have been comparatively quieter but follow the Model Code, so the same competence obligations apply.
Quebec's Barreau has its own Code of Professional Conduct. The bilingual content parity rules mean if you're using AI to translate, you own the translation's accuracy.
Several courts (Federal Court, Manitoba Court of King's Bench, and Yukon Supreme Court among them) have issued practice directions requiring disclosure when AI is used in preparing materials submitted to the court.

The FLSC's 2024 Model Code update (available at flsc.ca) is the plain-English baseline. If you haven't read the current version, that's the hour of homework to do this week.

For firms in regulated provinces (all of them, basically), the practical takeaway is simple. You are responsible for every word in every document that leaves your firm. AI is a tool. A hammer doesn't get blamed for a bent nail.

The Actual Cost of Getting This Wrong

Let me show the math, because I think when partners see the numbers they stop treating this as abstract.

Assume you're a 5-lawyer personal injury firm in Toronto. Your Google Ads CPC for "personal injury lawyer toronto" related terms is conservatively CA$40 per click (DataForSEO benchmark data shows "seo marketing for law firms" and adjacent keywords at CA$39.65 CPC for Google Canada, 2024, and personal-injury keywords typically run 2-3x higher in live auctions). Say you spend CA$6,000/month on ads. That's 150 clicks. Industry-average conversion from click to consultation request sits around 3-5%, so call it 6 consultations. Typical consultation-to-signed-retainer on PI work runs 20-30%.

So you're getting 1-2 signed files a month from that spend. Now one AI-hallucinated blog post goes live, the LSO flags it, you pay the compliance consultant CA$2,000 to review your whole site (that's the real number from a managing partner I spoke to in 2024), and you take the site down for two weeks to audit. You've just lost a month of intake. That's 1-2 files at average PI fee value. For most Ontario PI firms, that's somewhere between CA$15,000 and CA$80,000 in foregone revenue depending on case mix, plus the CA$2,000 cleanup, plus the reputational hit if another firm flags it to the Law Society.

The hallucination itself took 30 seconds to generate. Catching it would have taken 10 minutes.

That ratio, 10 minutes of verification vs. weeks of cleanup, is the whole argument for a process.

A Week-by-Week Process to Stop Hallucinations Before They Ship

This is the part that actually matters. Policy memos don't catch hallucinations. Process does. Here's what I'd implement in the first 60 days if I were running risk at a 2-15 lawyer firm.

Week 1: Audit what's already out there.

Pull every page of your website and every blog post published in the last 18 months.
For each piece with a legal claim, case citation, or statute reference, verify it. CanLII is free. Use it.
Flag anything that can't be verified within 5 minutes. Take those pieces down or rewrite.
Document who wrote each piece and whether AI was involved. You're building an audit trail.

Week 2: Write your firm's AI-use guidelines.

Not a 40-page policy. A one-page document. Who can use what tools, for what tasks, with what verification steps.
Explicitly ban: unverified case citations in any client-facing or court-facing document; AI-generated testimonials (banned under Ontario Rule 4.2-1 commentary regardless); AI drafting of any document that states a legal conclusion without a lawyer's sign-off.
We've got sample language in our firm AI policy guide if you want a starting point.

Week 3: Set up the verification checklist.

Every AI-assisted document (memo, letter, blog post, web page) passes through a two-person check before it's sent or published.
Person 1 (the drafter) verifies every cite and every statute reference against the primary source.
Person 2 (the reviewer) spot-checks at least 25% of citations and reads for factual claims that sound too specific.
Sign off in writing. Email thread. Matter notes. Whatever works. The audit trail is the whole point.

Week 4: Train the team.

45-minute session. Show real examples of hallucinations. Pull Zhang v. Chen. Pull the New York Air Canada case. Show what the fabricated cites looked like and how obvious they became under a 30-second CanLII search.
Run a live demo. Ask ChatGPT for "five Ontario Court of Appeal cases on [niche issue]." Verify each one on CanLII with the team watching.
The shock value of seeing a fake cite in real time is worth more than any memo.

Month 2, Weeks 5-8: Operationalize.

Every intake chatbot output gets a human review before it hits a matter file.
Every marketing piece with a claim or citation goes through the Week 3 checklist.
Every engagement letter for new matters should include language on AI use. We break down the provincial variations in our engagement letter AI language guide.
Quarterly review. Pull a sample of AI-assisted work. Audit it.

That's the process. It's not fancy. It's not a software purchase. It's a discipline.

Three Safeguards Worth More Than Any Tool

If you do nothing else, do these three things.

1. Never trust a citation you haven't clicked. CanLII, Lexis, Westlaw, the actual text. If you haven't seen the case, you haven't verified it. This one rule would have prevented every publicised AI-sanction case in Canadian and US courts over the past two years.

2. Assume the AI is wrong about jurisdiction. ChatGPT was trained on a lot of US law. It will cheerfully cite a California case as if it applies in Saskatchewan. The moment you see "Cal.", "S.D.N.Y.", or a Reporter abbreviation you don't recognize, stop and verify.

3. Keep AI out of anything sworn. Affidavits, pleadings, anything that goes to court, anything a client signs as true. Use AI for drafting bones if you want. But the final verification is 100% human. No exceptions.

If you're also wrestling with how to talk about AI in your marketing (without tripping compliance), we covered that decision in should you advertise AI-powered legal services.

FAQ: What Partners Keep Asking Me

Do we have to disclose AI use to clients? Depends on your province and what you're using it for. Several court practice directions require disclosure to the court. Many clients now ask as part of their own risk management. My rough rule: if you'd be embarrassed for the client to find out, you need to disclose. The safer practice is disclosure in the engagement letter.

What about Lexis+ AI and Westlaw Precision, are those safe? Safer. They're retrieval-augmented, meaning they pull from verified legal databases and cite the actual source. But "safer" isn't "safe." They still hallucinate, just less often and usually in summarization rather than citation. Verify everything.

Can we use ChatGPT for blog posts and marketing if we verify? Yes, with the Week 3 checklist process. The bigger risk with marketing content isn't fabricated cases, it's Law Society compliance on claims and testimonials. An AI will happily generate a testimonial. Ontario bans those outright.

What if an associate hallucinates a cite and files it without my knowing? The LSO and most law societies have been clear: supervision is the partner's duty. "I didn't know" isn't a defence. That's why the two-person verification checklist exists. It's not about trust. It's about evidence that you had a system.

Is there software that catches hallucinations automatically? Some. None are reliable enough to replace human verification as of 2026. Tools like Harvey, CoCounsel, and a few others claim detection, but the current state of the art still assumes a lawyer is checking the output. Treat any "AI-checks-AI" pitch with skepticism.