Written by: Wiktor Stefański, Head of People & Operations, Digital Colliers
The number that should be pinned above every legal drafting workstation right now: cases involving AI-fabricated citations rose from 87 to over 1,300 in the eleven months tracked in 2024. That is Stanford's Damien Charlotin count, and it is almost certainly an undercount, because it only captures the fabrications that got caught and made it into a written decision.
Read that curve as a leading indicator. Every one of those 1,300 was a filing that passed through some review process and still went out the door with a citation that pointed to nothing. The question for your firm is not whether your associates use AI. Most of them already do. The question is where in your workflow the fabrication check actually sits, and whether it sits there by design or by accident.
The check cannot live in the associate's head
The default state at most firms right now is that verification is implicit. The associate drafts with an AI assistant, glances at the citations, maybe pulls one or two, and passes it up. The partner assumes the associate checked. The associate assumes the tool got it right. Nobody owns the verification step, so nobody does it properly.
This breaks for three reasons:
- Fabricated citations often look correct. Right court, plausible party names, a reporter volume that exists. You cannot spot them by reading.
- Associates under billing pressure will not manually verify twenty citations. The ABA's Formal Opinion 512 already told us lawyers cannot bill hours that AI saved, which means the economic incentive is to move faster, not slower.
- Partners reviewing drafts are pattern-matching on argument quality, not running Shepard's on every cite.
If the check depends on a human remembering to run it every time, the check will fail. The 1,300 cases are the evidence.
Where the check has to sit
The verification step needs to be a gate, not a habit. It should run automatically between drafting and any partner review, and it should refuse to let a document proceed until every citation resolves against a real database record.
Concretely, that means three things wired together:
- The drafting tool (whatever your associates use, whether it is a legal-specific product or general-purpose) has to emit citations in a structured, machine-readable form, not just prose.
- A validation layer sits between the draft and the review queue. It hits Westlaw, Lexis, Bailii, CanLII, or whatever your primary citation database is, and confirms each cite resolves and says what the draft claims it says.
- The result gets logged into your case management system against the matter, so there is a durable record of what was verified, when, and by which version of which tool.
That third piece matters more than people expect. When a regulator or a court asks how the filing was produced, you want a timestamped record, not a reconstruction from memory. The UK Solicitors Regulation Authority put out its AI guidance in November 2023 and has been signalling for two years that firms are expected to have documented process control around AI use. That is not going to get looser.
The cost of leaving the gap open
The monetary cost of a sanction is the small part. The bigger costs stack behind it.
- One reported fabrication in your firm's name becomes the first search result for that partner for years.
- Insurance carriers are already pricing AI-related malpractice risk. Firms without documented verification workflows will pay more, or get non-renewed.
- Clients, particularly institutional ones, are starting to ask about AI use in engagement questionnaires. "We rely on associate diligence" is not going to survive that question much longer.
- Under EU AI Act Article 50 transparency obligations, which apply from 2 August 2026, certain AI-generated outputs need to be identifiable as such. Legal filings produced by firms operating in the EU will get pulled into that conversation whether the drafters want to be or not.
The firms that will look good in 2027 are the ones that treated citation verification as a plumbing problem in 2025. They bought or built the middleware. They wired the drafting tool to the citation database to the case management system. They made the check impossible to skip.
What to actually do this quarter
If you are reading this and you know your firm has the gap, the useful first move is small and diagnostic. Pick ten recent filings that involved AI-assisted drafting. Have someone verify every citation from scratch. Count the fabrications, the misquotes, and the cites that exist but do not support the proposition.
Whatever number you get is your baseline. If it is zero, your process is already working and you should document why. If it is not zero, you now know the size of the gap, and you can decide what to spend to close it before a court decides for you.