Written by: Wiktor Stefański, Head of People & Operations, Digital Colliers
The jump from 87 court cases with AI-fabricated citations to over 1,300 in eleven months isn't really a story about hallucinations. It's a story about workflow. Somewhere in each of those matters, a draft moved from a language model to a filing without a verification step that would've caught the problem. That gap is the actual signal, and it tells you exactly where your firm's process needs work.
The fabrication trend is a workflow signal, not an AI problem
LLMs invent citations. That's known, it's priced in, and every major bar body has said so. What the fabrication surge tells us is that the pace of adoption has run ahead of the pace of process design. Lawyers are drafting with AI on the side, then rushing to file. The tool didn't fail. The workflow around it did.
The UK's SRA published its AI guidance back in November 2023, and the ABA followed with Formal Opinion 512 in 2024. Both point the same direction. You can use these tools, but competence, supervision, and candour to the court still sit with the lawyer. The regulators aren't going to catch every filing. Your process has to.
What verification at scale actually looks like
When we're talking about a firm with hundreds of matters open at once, verification can't mean a partner squinting at a bibliography on a Friday night. It has to be a checkpoint that runs whether anyone remembers to trigger it. The teams getting this right tend to do a few things:
- Every AI-assisted draft passes through an automated citation checker that resolves each cite against a real database (Westlaw, Lexis, BAILII, CourtListener) before it leaves the associate's screen.
- The check produces a log. Which model, which prompt, which cites, which were verified, which were flagged. That log lives with the matter file.
- Flagged cites can't be silently overridden. Someone signs off, and that sign-off is recorded.
None of that is exotic engineering. It's the same discipline finance teams built around spreadsheet errors a decade ago. The firms treating verification as a system rather than a habit are the ones that won't show up on the Stanford tracker next year.
Drafting, matter management, and citation check need to be one loop
Right now, in most firms, these are three different tools that don't talk to each other. The associate drafts in Word with a Copilot-style assistant. Time and billing happens in Aderant or Elite. Citation checking is a browser tab. That gap is where fabricated cites survive.
The integration doesn't have to be a giant platform rebuild. What it does have to do is make the audit trail automatic. If an AI draft touches a matter, the matter record should know. If a cite gets verified, the matter record should know. And this matters for billing too. ABA Opinion 512 is explicit that you can't bill hours the AI actually saved. Given that only about three of a lawyer's eight hours get billed on average, the firms that can accurately show AI-assisted time versus lawyer-verified time are going to have a cleaner story for clients and regulators alike.
What a managing partner can actually do in the next quarter
You don't need a two-year transformation programme. You need three things visible by the next partners' meeting.
- A written AI-use policy that names the tools, the permitted uses, and the mandatory verification step. Short. Read by everyone.
- A citation verification workflow that runs automatically on any filing drafted with AI assistance, with a log attached to the matter.
- A billing review that reconciles AI-assisted time against billed time, so you're not on the wrong side of Opinion 512 or its UK equivalents.
And keep an eye on the calendar. EU AI Act transparency obligations under Article 50 kick in on 2 August 2026, with high-risk obligations following on 2 December 2027. For firms advising EU clients or operating in the EU, those dates aren't distant.
The left-behind risk is quieter than you think
The firms that end up behind on this won't lose in one dramatic filing. They'll lose the way firms always lose ground. A client asks how AI is used on their matter. The answer is vague. The client moves work to a firm with a cleaner answer. Meanwhile, the firms that built the verification loop early are quoting on the same work with an audit trail in hand. That's the gap that opens up in 2026 and 2027, and it opens up quietly.