Jump to section
A lawyer in another province filed a factum built on cases that did not exist. The citations looked perfect. The style of cause was plausible, the reporter series was real, the pinpoints were formatted like any other pinpoint. The only problem was that a chatbot had made all of it up, and nobody had opened a single one of those decisions before the document went to court. That is the whole story of AI in practice right now, compressed into one embarrassing filing: the tool is fast, the tool is confident, and the tool does not care whether it is right.
I am not writing this to scold anyone off the technology. I use these tools most days. But I want to make an argument that I think too many firms are dodging: when AI touches your work product, you are still the author. Not a supervisor, not an editor, not a curator. The author. That single idea should reshape how you review everything a model hands you.
The Model Is a Draftsperson, Not a Lawyer
Think about how you already treat a first-year drafting a memo. You do not sign it and send it because the format is tidy. You read it, you check the authorities, you push on the reasoning, and if it holds up, it becomes yours. A language model deserves less trust than that first-year, because the model cannot tell you which sentence it is unsure about, has no idea whether its confidence is warranted, has no professional obligation, no insurer, and no career to lose.
The law society does not have a rule that says "AI is fine as long as you check it," because it does not need one. Your existing duties already answer the question. Competence, candour to the tribunal, supervision of work done on your file: these obligations attach to the finished document no matter what produced the draft. The model is a draftsperson you happen to be able to summon at any hour. It is not a colleague, and it is certainly not counsel of record.
Note. The right mental model is delegation, not automation. You would never let an unsupervised clerk file a factum. Treat model output the same way, and most of the hard questions answer themselves.
Where Confident Wrong Answers Hide
The dangerous errors are not the obvious ones. If a model tells you the limitation period is "roughly a while," you will catch it. The trouble is the output that reads exactly like good work. Hallucinated citations are the famous example, but they are only the most visible failure. Underneath them sit quieter problems: a rule stated correctly but for the wrong jurisdiction, a case summarized accurately except for the part that actually mattered, a paraphrase that flips a holding while keeping every surrounding sentence true.
These errors hide precisely because the surrounding text is fluent. Polished prose can make a wrong answer look as credible as a correct one, and the more capable the writing sounds, the harder it is to stay critical. A confident wrong answer does not announce itself.
The more fluent the output, the more carefully you need to check it, because producing fluent text is exactly what these tools do well, and verifying factual accuracy is exactly what they do not. A working rule I hold to
Checking Facts, Citations, and Reasoning Separately
Here is the practical heart of it. Do not "review" AI output as one undifferentiated pass. That is how errors slip through, because a single read blends three very different checks into a vague sense that it "looks fine." Split the job.
- Facts. Every factual assertion about your matter, your client, dates, amounts, parties, gets checked against the record, not against the model's own summary of the record.
- Citations. Open every case. Actually open it. Confirm the decision exists, the citation is correct, the proposition is really in there, and the pinpoint lands where the text says it does. If you cannot open it, you cannot cite it.
- Reasoning. Read the argument as an adversary would. Does the logic actually connect the authority to the conclusion, or does it just sound like it does?
Three passes, three different mindsets. The facts pass is clerical and careful. The citations pass is mechanical and unforgiving. The reasoning pass is where your judgment as a lawyer does the work no tool can do. If you are researching genuinely novel questions, our note on doing legal tech well and the broader knowledge base both reinforce the same point: verification is not a formality you can skip when the deadline is tight.
Building a Review Ritual Into Your Workflow
Discipline that depends only on remembering will fail on your worst day. The filing that goes wrong is almost never the one you had time for. It is the 11 p.m. draft before a hearing, when you are tired and the output looks done. So the review has to be a built-in step, not a matter of remembering.
A few habits that survive contact with a real practice:
- Keep the prompt and the raw output. If a draft ever comes into question, you want to reconstruct what the tool actually said versus what you changed.
- Never let model text pass straight into a filed document without a human editing pass that visibly changes something. Untouched output is a red flag to your future self.
- Log which authorities you personally opened. A simple checklist beside your citations list is enough.
- Decide in advance what work the tool is allowed to touch, and write it down where your team can see it.
That last point is a policy question, not a personal one. If you have staff or juniors using these tools, the standard cannot live in your head. Our companion pieces on a small firm AI policy that works and on keeping automation under human control go deeper on writing that down. If any of this touches client data, the confidentiality questions in our cloud tools piece deserve a read before you paste anything sensitive into a prompt. The tools you already use to run the firm, from matter management to your billing, are the natural place to record which drafts had AI involvement and who verified them.
Warn. "The AI told me" is not a defence to the tribunal, to your client, or to your insurer. The obligation to verify is yours, and it does not transfer to the vendor no matter what their marketing says.
What to Tell the Client and the Court
Candour is where the opinion gets uncomfortable, so let me be direct. You do not owe the court a disclosure every time a tool helped you format a heading, any more than you disclose which word processor you used. But you absolutely owe the court accurate authorities and honest submissions, and if a court in your jurisdiction has issued a practice direction on AI use, you follow it to the letter and confirm the current version before you file.
With clients, the calculus is different and, I would argue, simpler. Be honest that you use modern tools, and be honest that you verify everything they produce. Most clients do not want to hear that their lawyer refuses to use anything invented after 1995. They want to hear that you are efficient and that nothing goes out the door unchecked. That is a comfortable, truthful thing to say, and it happens to be exactly the standard your licensing obligations already demand.
The Takeaway
In practice the verify step is the whole job. The model is a fast, tireless draftsperson that can produce a wrong answer with the same confidence as a correct one, and your name is the one on the document. Read the facts, open the cases, test the reasoning, and make those checks so routine that skipping them feels wrong. Do that, and AI becomes what it should be: a tool that saves time on the drafting while you keep responsibility for the substance. The technology is not the risk. Failing to check the output is.
