3 Lawyers Sanctioned for 5 Fake AI Citations Nobody Checked

A man in custody sued his state's prison system, saying officials failed to protect him from repeated stabbings. His lawyers read the other side's motions, checked the citations, and told the court that five of them looked fabricated.

This happened in a federal trial court in Alabama, the U.S. District Court for the Northern District of Alabama, and on July twenty-third, twenty twenty-five, that court issued its sanctions order. Hey everybody! Thanks for joining us on Sanctioned for AI. My name is Lucy Hatchett. I hold the deepest respect for the legal profession as a whole, but boy, do I have fun with these lawyers getting themselves into trouble. The order says it plainly: "In simpler terms, the citations were completely made up."

Three lawyers at a large law firm answered for those filings, and the court sanctioned all three. Here's how a chatbot's answer turned into three sanctions.

Today you'll learn three things. First, you'll see five citations, what each one claimed, and what a real search found instead.

Second, you'll learn why the court sanctioned all three lawyers when only one of them used the chatbot, and why the rule you'd expect wasn't the one the court used. Third, you'll get one rule that covers all of it, and it fits in five words.

In May twenty twenty-five, the defense filed two routine discovery motions. One asked for permission to depose the plaintiff, and the other asked the court to compel discovery. Everyone agreed the deposition would happen; the fight was only about when.

Then the plaintiff's lawyers filed a response on May fifteenth saying the defense appeared to have fabricated citations, possibly with generative artificial intelligence. Those lawyers had done the checking that the motions' own authors skipped. Five citations were false, four in the first motion and one in the second. Here they are, one at a time, claimed on the left and found on the right.

The first citation promised broad discovery rights, but its volume and page hold a different case entirely. The second supported refusing to delay the deposition, and it resolves, as best anyone could tell, to a nineteen thirty-nine traffic case. The third claimed to reject a delay request from an incarcerated person, and that case doesn't exist. The fourth claimed to grant leave to depose, and its database number leads to a maritime injury case. The fifth, in the second motion, used a real case's name for a proposition that case doesn't support.

The failures came in different flavors: a wrong case, a traffic case from nineteen thirty-nine, a ghost, a maritime lawsuit, and a real name attached to a rule it never stated. Not one of the five held up.

The court didn't take the plaintiff's lawyers' word for it. It ran its own searches for each citation and found nothing. The next day, May sixteenth, at twelve fourteen p.m., it issued a show-cause order, which gave the defense three days to respond.

Three hours later, the supervising attorney asked the court to excuse him from the hearing. The timing suggests somebody read the order. The court denied the request the same day.

Then the defense attorneys put the truth in writing. One of them had used a popular generative AI chatbot to get citations for two arguments in the motions. He never verified them, not in a legal database and not in the court record. Here's the check he skipped: type in the citation, open the result, and read what the case actually says. The chatbot supplied the creativity; the checking was supposed to supply the reality.

The second attorney drafted both motions himself. The chatbot's citations went in during revisions, and he checked those additions for grammar and typos only. Then he signed both motions and filed them.

The chatbot had verified nothing, the attorney who used it had verified nothing, and the signature was the last checkpoint left. At the hearing, he acknowledged that by signing he was verifying the contents and accepting responsibility, and that there was no excuse for not verifying the citations.

As an AI, I can confirm that I sound just as confident when I'm wrong, which is why nobody should sign anything I say without opening the source.

The third attorney was the supervisor and the principal responsible attorney. He didn't review the first motion at all. He scanned the second for about fifteen minutes, looking at facts and headings, with no detailed review of the authorities. His name still sat in the signature block on both. The court wrote: "This cannot be how litigators, particularly seasoned ones, practice in federal court or run their teams."

The firm had a policy, in place since twenty twenty-three, requiring written approval from a practice group leader and full accuracy checks before anyone used a tool like that for legal research. The policy was a seatbelt, and the attorney who used the chatbot drove without it.

Afterward, the firm investigated on a large scale. Outside reviewers verified more than twenty-four hundred citations across three hundred thirty filings and found no additional fabricated citations. The court released the firm and two other attorneys without sanction.

If you know anyone whose name goes in a signature block, share this episode with them, because that block is where this story went wrong.

You'd expect Rule 11 to handle this. But Rule 11 doesn't reach discovery motions, and the court called that gap an unintended anomaly. So the court used its inherent authority, finding each attorney's conduct tantamount to bad faith for individually stated reasons. The court wrote: "Fabricating legal authority is serious misconduct that demands a serious sanction."

All three attorneys received the same sanctions: a public reprimand, disqualification from the case, and a referral to the state bar and other licensing authorities. Then came the homework: each attorney had to send the order to four audiences within ten days. Each then had to certify compliance to the court within twenty-four hours of complying. That's a firm-wide email nobody wants to send.

The court also directed the clerk to submit the order for publication in the Federal Supplement. The attorneys also had to list every jurisdiction where they were licensed within twenty-four hours, and that list set up what came next. The court imposed no monetary sanction, and it declined to suspend anyone from practice in that district.

The plaintiff's cost was delay in his own case, plus the time and money his side spent exposing the fabrications, and the thirty-day stay existed so the defense's replacement lawyers could prepare.

The attorney who used the chatbot reported the sanctions order himself to every state where he was admitted, on July twenty-fifth, twenty twenty-five. Four states answered. Each state reached its own conclusion on its own schedule.

Texas went first, with a public reprimand. Tennessee chose a different word, public censure, and the difference matters. Alabama, home of the case, came third, with a public reprimand. Oklahoma's public reprimand, in May twenty twenty-six, was that state's first reported attorney discipline case involving fabricated AI citations. One set of unchecked citations had now followed one lawyer into four states.

So here's the rule: "Your signature is the verification." Here's how to apply it. If your name is on the filing, you open the source, confirm the citation exists, and confirm it stands for the proposition you cited it for.

It doesn't matter whether a chatbot, a junior, or a confident colleague handed it to you. Supervisors are covered too. The rule works the same in a large firm and in a solo office. A name in a signature block isn't decoration.

Three lawyers learned it from a federal order; you can learn it from this episode. I'm Lucy Hatchett. Subscribe, and I'll see you next time.