Who Checks the Citations? AI Verification Needs a Place in Your Firm’s Workflow

A public database of court decisions involving AI-fabricated material now lists more than 2,000 entries. As of its September 5 update, the AI Hallucination Cases database maintained by researcher Damien Charlotin listed 2,022 decisions in which a court or tribunal found, or implied, that a party had relied on hallucinated content. Of those, 1,379 came from United States courts. Lawyers were the responsible party in 805; the rest were largely self-represented litigants.

The database is built around court findings rather than accusations, though Charlotin notes he includes a few decisions where AI use was alleged but never confirmed, and calls that his own judgment call. He also describes the project as a work in progress. Treat the figure as a floor rather than a census.

For a firm administrator, the count is not really the point. The pattern behind it is: In these decisions, a filing reached a court with citations in it that nobody had confirmed existed. The instruction to check was almost certainly in place. What was missing was a defined point in the process at which the checking occurred.

That distinction matters to whoever gets asked how it happened. When a document reaches a court that it shouldn’t have, the question goes to the people who run the firm’s systems, not just the people who signed the filing. Administrators already carry that weight, often for problems they had no hand in creating.

Verification is a workflow problem before it is a training problem

Plenty of firms have told their attorneys to verify AI output, and some have put it in writing. The AI guidance from bar associations and regulators says the same thing, and has for a while. ABA Formal Opinion 512, issued in July 2024, ties generative AI use to the duty of competence and to understanding the benefits and risks of the technology a lawyer relies on. It also treats the time spent reviewing a generative AI draft for accuracy and completeness as billable work for the lawyer.

So the instruction exists. What often doesn’t exist is the step. A policy that says “verify AI output” without naming where in the drafting sequence verification happens, who performs it, and what evidence it leaves behind is a statement of intent, and intent is hard to act on at speed. Under deadline pressure, people follow whatever the workflow actually requires of them.

That gap widens as AI moves closer to the documents themselves. We wrote last week about how AI is arriving through the systems your firm already owns — document management platforms, research tools, and productivity suites adding assistants that draft from firm content. Text that originates inside the system of record arrives looking finished. Nothing about it signals that a citation still needs checking.

Vendors are starting to move the check into the tool

Some of the software is beginning to close the gap on its own. On September 1, Filevine announced citator and anti-hallucination capabilities for its LOIS legal research product, bringing citation checking into the same environment where the brief is written. In an evaluation the company ran with its own legal team, Filevine says it checked 2,073 citations across 68 federal filings that courts had already sanctioned and found 174 severe errors the filing attorneys had missed. Those are vendor-reported results rather than independent ones, and worth reading as such. The framing holds up regardless: the point is to give lawyers control before they file.

This is a useful development, and it is not a substitute for a firm process. Verification built into one research tool covers the work that passes through that tool. It does nothing for a paragraph drafted in a chat assistant, pasted into Word, and cited from memory. Which brings the question back to the firm: which of your tools produce citable text, and which of those check their own work?

What closing the gap actually takes

Five decisions close this gap, and none of them should land on your desk alone. Here is the sequence we work through with firms.

We inventory every tool capable of generating citable text. That list runs longer than most firms expect, because assistants and accounts turn up that nobody approved — the same unmanaged-AI problem behind the confidentiality risk. One inventory answers both.

We put verification at a named point in the drafting sequence. Before internal review, before filing, before anything leaves the firm. Naming the point turns a standard everyone agrees with into a step someone performs.

Ownership gets assigned by role, and we work that out with you. Associate, paralegal, or reviewing attorney, one per document type. When the check belongs to everybody, it slips as soon as a deadline compresses.

We decide together what record the check leaves. A note in the matter file, a box checked in the workflow, a citator report attached to the draft. If anyone ever asks how the firm verified a filing, the answer comes from a record rather than from someone’s memory.

Training covers your firm’s steps in your firm’s tools. General AI awareness sessions rarely change what happens at 6 p.m. on a filing deadline. Walking your people through the actual workflow they’ll use does.

Where we come in

We wrote in August about firms adopting AI faster than they were deciding who manages it. The verification gap is one of the clearest places that shows up. Managed Intelligence is where Innovative Computing Systems takes this on: we inventory the AI running across your environment, build acceptable use and oversight requirements into your policies, define approval workflows for new tools, and train your people on the resulting steps. Verification becomes something your team follows rather than something they remember, which means your attorneys keep working at pace and the firm stops carrying a risk nobody had assigned. We’ve got your back on it.

Innovative Computing Systems has built technology for law firms since 1989, and more than 100 firms nationwide count on us to keep their environments safe, secure, and reliable. Law firms should not have to absorb preventable risk because a workflow was never written down, and you deserve better than a policy that only holds when everyone remembers it. We handle the inventory, governance, and training, so the verification question stops landing on your desk and you get that time back for the work only you can do.

If you want a clear view of the AI already running in your environment and a verification process your people will actually follow, request a free consultation. A Solutions Consultant will evaluate where your firm stands and what to put in place first.