It usually starts the same way.

You’re up late, closer to midnight than you’d like to admit. Finishing a factum that felt manageable at noon and now feels like it’s grown sentient. The argument is there. The structure holds. You need a few more authorities to cleanly land the point.

You find them. Or something finds them for you.

They look right. The citations are clean. The reasoning tracks.

You paste them in. You move on.

And nothing in that moment tells you that you’ve just crossed a line, not stylistically, not rhetorically, but epistemically.

Because the problem is no longer just what the law says.

It’s whether what you’ve just put before the court is real.

1. The Quiet Shift in Litigation Practice

We tend to frame AI risk as technological. That misses the point. The real shift is this: we are now surrounded by tools that produce answers without guarantees of truth. And in litigation, that creates a new kind of professional exposure. Not because the rules have changed, but because the failure modes have. The case law has caught up quickly. Ko v Li, 2025 ONSC 2965 (CanLII) is the obvious example: a factum, plausible on its face, built on authorities that did not exist at all.

But even that framing is too narrow. What matters is not the tool. It is the pattern. Because not all citation problems are the same, and courts know the difference.

2. Three Ways to Mislead a Court (Without Meaning To)

Let’s start where most of us have been.

• Selective Omission: The Incomplete Truth.

Imagine this. You’re arguing a duty of care. You find a strong appellate case supporting your position. It’s clean, well-reasoned, helpful. You cite it. What you don’t emphasize, is that on the facts, the court ultimately declined to impose liability.

You haven’t lied. Not quite. But if opposing counsel stands up and says, “my friend hasn’t taken you to paragraph 78 …,” that small omission suddenly feels much larger. Courts understand advocacy. What they are attentive to is whether the omission changes the shape of the law the court thinks it is applying.

This is the most human category of error. It lives closest to judgment. And it is often where credibility begins to erode, not because of what was said, but because of what wasn’t.

• Mischaracterization: The Distorted Truth.

Now move one step further. You cite the case. You’ve read it. You believe it helps you. But in tightening the argument, compressing, streamlining, sharpening, you push the principle further than the judgment actually does.

A qualified statement becomes absolute. A contextual analysis becomes a rule. If you’ve been in court long enough, you’ve seen the moment this lands badly. The judge leans forward: “where exactly does the case say that?”

That question is doing more work than it seems. It is not just about the case. It is about whether the court can trust you as a narrator of the law.

Mischaracterization is where the issue shifts from persuasion to reliability. And the dividing line becomes uncomfortable: Did you get this wrong, or did you need it to be right?

• Hallucination: The Invented Truth.

Now imagine something different. You cite a case. The name is plausible. The citation looks correct. The quoted passage is elegant. Opposing counsel can’t find it. The judge can’t find it. You can’t find it. Because it was never there.

This is the category that has brought AI into sharp focus. Not because hallucinations are common, they’re not, but because they are convincing. They don’t look like errors. They look like shortcuts.

But from the court’s perspective, the characterization is stark: You have placed something before the court that does not exist. At that point, it no longer matters whether the source was an AI tool, a research memo or a junior who misunderstood the assignment. The question is simpler and harsher: How did this get signed, filed, and relied upon without being known to be true?

3. Where Everything Turns: The Candour Moment

Up to this point, all of these errors, even the serious ones, live in the realm of professional competence. They can be fixed. The factum is withdrawn. The record is corrected. An explanation is given. And then there is a second moment. A quieter one. The judge asks: “How did this happen?”

This is where cases like Ko v Li, 2025 ONSC 2965 (CanLII) become something else entirely. Because the outcome no longer turns on the error. It turns on the answer.

• Was the explanation complete?
• Was responsibility accepted?
• Was the court told, fully and directly, what occurred?
• Or was there hesitation? Deflection? A narrative that softened, shifted, or redistributed responsibility?

That is the hinge.

The moment candour falters, the problem is no longer accuracy. It is integrity. And that is when a correctable mistake becomes something that engages the court’s contempt jurisdiction.

It’s worth pausing to recognize something simple but impactful: we’re all still learning this. Legal AI arrived faster than our habits could adjust, and most of us have experimented with it in the middle of real deadlines, real pressure, and real expectations from clients and courts. It’s not surprising that the lines between efficient use, over-reliance, and risk aren’t always clear in the moment. The goal here isn’t to judge those missteps harshly, it’s to understand how they happen, and more importantly, how we can build practices that keep us anchored in accuracy, judgment, and candour going forward.

4. A Draft Decision Tree for the Real World

Strip away the doctrine, and the workflow becomes surprisingly clear. It’s less about AI than about discipline.

• Step 1: Does the case exist? If you haven’t pulled it from a reliable source and read it, you don’t know that it does.
• Step 2: Does it say what you think it says? If the proposition feels cleaner than the judgment, it probably is.
• Step 3: What have you left out? If there is a paragraph you hope no one asks about, assume someone will.
• And over all of it, Step 4: If something goes wrong, what will you say? Because you will be asked.

‍5. Five Questions Before You Hit “File” or “Submit”

The discipline can be reduced even further. Before a factum is filed, or before you stand up in court, there are five questions that should feel almost automatic:

1. Have I opened and read the case(s) I’ve cited?
2. Does each authority I rely on exist and can it be independently verified?
3. Does the proposition I attribute to each case accurately reflect its ratio?
4. Have I disclosed any material adverse authority?
5. Would I be comfortable defending each citation, its existence, accuracy, and fairness, under scrutiny in open court?

If any answer is “I think so,” or “probably,” that is already a flag.

6. The Illusion We’re Losing

There is a subtle illusion that AI has disturbed. For a long time, lawyers have worked with layers: research prepared by others, authorities inherited from earlier drafts, arguments built on prior work. We have always relied, to some degree, on the assumption that someone upstream has verified the foundation.

AI accelerates this dynamic. It produces polished, coherent, complete-seeming work product instantly. What it removes is friction, the pause where verification used to happen. And so the role of counsel becomes clearer, not looser: lawyers are the last point at which plausibility must become truth.

At its core, this is not a story about technology. It is a story about a professional obligation that has become harder to perform, precisely because it looks easier.

The relevant question is no longer: “Does this look right?” It is: “Do I know this is true?” And that difference, small in phrasing, enormous in consequence, is what now separates strong advocacy from unreliable advocacy, error from misconduct, and, in the rare but real cases, mistake from contempt.

In that sense, nothing has changed. And everything has.