The Opinions You Cite May Have Had Help

For two years the AI verification burden in litigation has pointed one direction. A lawyer files a brief, and that lawyer answers for every citation in it. Courts have been enforcing that with rising force and shrinking patience.

An analysis published August 24 turns the same question toward the other side of the bench.

What was actually measured

Josh Morrow, a partner at Lehotsky Cohn, ran roughly 2,250 published opinions from the regional federal courts of appeals, covering January through early August 2026, through Pangram, a commercial AI-detection tool. More than 50 came back showing signs of AI authorship. The detected percentages ran from under 1% to over 50%, clustering toward the low end. As a control he ran more than 300 opinions from January 2022, before these tools were in general use. Those returned nothing.

Morrow is careful about what this establishes, and we are going to be equally careful, because the caveats are most of the story. He writes that the evidence "is not conclusive" and that the detection is "neither foolproof nor comprehensive." Detection tools produce false positives on formal, heavily structured prose, and a published appellate opinion is about as formal and structured as written English gets. In the other direction, a judge who uses a model to research an issue or rough out a background section may leave no detectable trace in the filed document, so the method misses cases too.

The finding is not that federal appellate courts are drafting opinions with AI. The finding is that a 2022 control set registered zero, a 2026 set registered more than fifty, and the gap between those two numbers is large enough to deserve an answer from someone who actually knows.

We would also note who ran it. This is a practicing appellate lawyer with a detection tool and a weekend, not a judicial conference study. That is a real limitation. It is also how most of the useful questions in this area have surfaced so far.

Why this reads differently inside a firm

Take the finding at its weakest and it still changes something.

Your associates verify AI output because the tool can fabricate. That discipline was built on the premise that the underlying corpus of law is human-written and reliable, and the risk sits entirely in the drafting layer your firm controls. If some portion of the opinions in that corpus were themselves partly machine-drafted, the premise gets softer. Not broken. Softer.

Practically, an opinion is the law whether a clerk wrote it, a judge wrote it, or a model helped. It is signed, it is published, and it binds. Nobody should be advising clients that a circuit opinion is less authoritative because a detection tool flagged its prose. That argument will not go anywhere good.

What does change is the reliance question inside your own work. A published opinion that summarizes a prior case in a way that is subtly off, or characterizes a record in language that reads clean but drifts from the underlying facts, propagates. Your brief cites the summary. Opposing counsel cites the summary. The next opinion cites yours.

That failure mode is not new. Clerks have always written most of the prose in most opinions, and summaries have always drifted. What is new is the volume at which the drift can be produced, and the fact that it now reads well while doing it.

The other direction is already expensive

While the question about the bench stays open, the question about the bar is settled and getting costlier.

On July 28 the Illinois Appellate Court, First District, sanctioned an attorney $15,000 and referred him to the state disciplinary commission over briefs containing ten false citations. The court calculated the figure at $1,500 per fabricated citation and said so on the record, explaining that "courts have no choice but to increase fines for AI-hallucinated citations until those fines have a significant deterrent effect."

Read the mechanics of that sentence. The court did not pick a round number. It built a per-unit rate and stated it was setting the rate above prior cases because prior cases had not worked. That is a schedule, and schedules go up.

The attorney said he had used ChatGPT and cross-referenced the results in LexisNexis. His response brief, filed after opposing counsel flagged the problem, contained further errors.

What we would do

Separate the two questions and answer only one of them. Your firm's exposure is your filings. The authorship of published opinions is interesting, unresolved, and not something you control. Do not let the second question become a reason to relitigate the first.

Verify against the record, not against the summary. If a proposition is load-bearing in a brief, pull the underlying case and read what it actually held, rather than citing a later opinion's characterization of it. This is old-fashioned advice that stopped being routine when full-text search made the summary easy to find. It is the specific practice that survives whatever the answer about AI drafting turns out to be.

Price the verification, then decide who does it. A $1,500-per-citation rate makes the cost of a missed check calculable for the first time. Firms have been treating citation verification as something that happens somewhere in the review chain. Name the person.

Do not run detection tools on judicial opinions and put the results in a filing. We expect somebody to try this within the year. Given the false-positive rate on formal prose, it is a losing motion and a bad look, and it will draw the court's attention to your own drafting process.

Where this goes

The honest position right now is that one careful practitioner found something worth checking and said out loud that his method has limits. Courts will eventually address this, probably through disclosure practice rather than prohibition, the same way they arrived at standing orders for attorney filings.

Until then the useful move is not to have an opinion about what judges are doing. It is to make sure that when the question gets asked properly, your own filings are not part of the story.

If your firm is working out where AI sits in its research and drafting workflow, and what the verification chain actually looks like on paper, that is a conversation we are glad to have.