At a September 4 hearing before the Second Circuit, an attorney had to explain errors in a brief filed in an $82 million subway-injury appeal. The brief cited real cases but presented paraphrases as direct quotations.

According to the ABA Journal’s report, attorney Elliot Shields acknowledged using artificial intelligence to write the brief and missing mistakes. He had notified the court about the errors before oral argument. At the hearing, he said, “We were rushing, and we were sloppy,” and explained that he believed he had transposed notes and placed them in quotation marks.

The panel gave him one week to file a corrected brief.

The reporting does not establish that AI generated every inaccurate quotation. It does show why a citation check that stops after confirming the case exists is incomplete.

A real case can still be cited incorrectly

Fabricated cases get most of the attention because they are obvious once someone searches for them.

A misquotation is harder to catch. The case name and reporter citation may be correct, and the language may sound like something the court would have written. A quick existence check passes even though the sentence between quotation marks never appeared in the decision.

An accurate quotation can also be used inaccurately when it comes from a party’s argument, a lower court, or a dissent rather than the controlling analysis.

This is narrower than the failure we covered when Sullivan & Cromwell corrected fabricated citations, pin cites, and parenthetical quotations. It is also easier to imagine slipping through a busy review. The source opens. The case looks familiar. The sentence fits the argument. Nothing appears obviously fake.

One checkbox can hide several different checks

Many filing workflows reduce citation review to one box: “Citations checked.” That makes it impossible to tell what the reviewer actually confirmed.

There are at least four separate decisions buried inside that box:

Does the authority exist?
Does the source contain the quoted language?
Does the cited passage support the proposition in context?
Is the relevant point still good law?

An existence check answers only the first question. A conventional citator focuses mainly on the fourth. Neither proves that the quotation is accurate or that the opinion supports the sentence in the brief.

The process also needs a named reviewer. A checklist, document status, or matter note should identify the lawyer responsible for the final quotation and authority review. “Someone reviewed it” is difficult to rely on when no one is named and a deadline is approaching.

New tools can help, but they are not the sign-off

Legal technology companies are building products for this exact gap. BriefCatch says its newly released RealityCheck feature compares quotations, holdings, dispositions, and propositions against cited cases. Filevine says its LOIS brief auditor flags altered quotations and authorities that do not support the proposition attributed to them.

Those are vendor claims, and both products may help a reviewer find a problem sooner. They can also produce an unverified or ambiguous result that a lawyer must resolve.

An automated checker works best as another control before the human sign-off. It should not generate the sentence, perform the only review, and effectively approve its own output.

Firms evaluating one of these tools should test it with examples they already understand. Include a fabricated case, an altered quote from a real case, a correct quote taken out of context, and a proposition supported only by a dissent. See what the product catches, what it misses, and what evidence it provides to the reviewer.

Make the check visible in the workflow

A policy that says “verify AI output” leaves too much room for interpretation. One attorney may confirm that the cited case exists. Another may check quotations but not later treatment. Under a deadline, both may assume someone else completed the final pass.

The firm’s filing process should answer:

Who verifies every direct quotation against the source?
Who confirms that each authority supports the proposition?
Where is that review recorded?
Can a filing move forward if the verification step is incomplete?

The same workflow should apply even when no one reports using AI. Lawyers have misquoted cases for much longer than generative AI has existed. AI raises the risk because it can produce plausible legal language quickly and without visible uncertainty, but making the control depend on self-reported AI use creates another way for the check to be skipped.

The Second Circuit hearing shows why citation verification must go beyond opening the case. A source can exist and a citation can resolve while the filing still misquotes the court. The workflow needs a named reviewer who checks the words against the source.

If your firm’s AI policy says “verify the output” but the filing workflow does not show who checks what, Artech can help turn the policy into a repeatable, documented process.