Joyce Barber v. Lawrence J. Morawa, MD was an ordinary medical malpractice case that turned into a small monument to the idea that you should never try to put out a fire with a bucket of the same thing that started it.

Barber lost at the trial level and was denied a new trial. Her lawyer, Ronnie Cromer Jr. of Southfield, took the case up to the Michigan Court of Appeals. That is a completely normal move. Losing parties appeal every day. What was not normal was the legal authority Cromer brought along for the ride.

Caught once

The appellate brief cited cases that do not exist. Not cases misread, not cases stretched past their holding, but citations to authority the court simply could not find, because it had never been real in the first place. This is the signature failure mode of a general-purpose language model asked to do legal research: it produces text that has the shape, rhythm, and confident formatting of a real citation, complete with a plausible reporter volume and page number, and none of the underlying reality.

Courts have a name for the lawyer's obligation here, and it predates ChatGPT by decades. Before you put a proposition in front of a judge with a citation attached, you are supposed to confirm the cited authority exists and actually supports the proposition. The Court of Appeals flagged the fabricated authority and gave Cromer the chance every court gives: explain yourself and fix it.

The correction that needed correcting

This is the part that lifts the story out of the now-crowded genre of "lawyer files fake cases." After being alerted, more than once, that his filings contained fabricated authority, Cromer filed a Notice of Correction roughly three weeks before oral argument. The document was supposed to scrub the bad citations and reassure the court that the grown-ups were back in charge.

Instead, it produced a fresh batch of bad authority. According to the panel, Cromer used AI to write the correction too. This time the cases he cited were real, but he attributed quotations and propositions to them that those cases never contained, including language that actually lives in a rule of evidence rather than in the case he pinned it to. He brought a mop made of the same spilled water.

It is hard to overstate how much this detail matters. A single batch of hallucinated cites is the kind of thing a sympathetic judge might chalk up to a one-time lapse in supervision. Doing it again, in the specific document whose entire job was to demonstrate that you had finally checked your work, removes the lapse defense. It shows the verification step was never added, even after the court personally pointed at the missing step and asked where it went.

At oral argument

By the time Cromer stood up for oral argument, the explanation was the one the legal profession has heard on a loop since 2023: he had used AI tools. The panel was unmoved. Its opinion described "repeated reliance on artificial intelligence without meaningful verification, despite having been alerted more than once that his filings contained fabricated authority." That sentence is doing a lot of quiet work. It is not punishing the use of AI. It is punishing the refusal to check the output after multiple warnings that the output was wrong.

What the court did about it

The panel did not settle for a stern paragraph and a reminder to do better. It marked the opinion FOR PUBLICATION, which means it becomes citable precedent and, more to the point, a permanent public record with Cromer's name attached. It concluded that the repeated submission of fabricated and unsupported authority violated the Michigan Court Rules and the duty of reasonable inquiry.

Then it reached for consequences with actual weight. The court found the appeal vexatious and remanded to the trial court to determine the real damages and expenses, including reasonable attorney fees, that the defendant incurred because of the appeal, payable by Cromer personally. Not by the client. By the lawyer. The panel also directed the clerk to forward the opinion to the Michigan Attorney Grievance Commission for possible investigation, which is the polite institutional phrasing for "the people who can suspend your license would like a word."

Why this one is worth filing away

There is a version of this story that is just another tally mark on the long list of attorneys sanctioned for trusting a text generator. What makes Barber v. Morawa stick is the structure of the failure. The legal system is built around second chances and the assumption that a professional, once told their work product is defective, will go verify it. Cromer got that chance in the most explicit way possible, responded with a correction, and the correction was defective in exactly the same way for exactly the same reason.

The lesson is not "AI is dangerous in law," though it can be. The lesson is narrower and older. A citation is a promise that you looked. When you hand a court a brief, you are vouching that the authority is real and says what you claim. A language model cannot make that promise for you, because it has no idea whether the case exists; it is producing the most probable next tokens, and a fake citation is often more probable-looking than a real one.

The fix is unglamorous and has not changed in decades. Open every cited case. Confirm it exists. Confirm it stands for the point. Confirm the jurisdiction. If that sounds tedious, congratulations, you have rediscovered the part of legal research that was always the actual job. The generator can draft. It cannot vouch. The signature on the filing, and now the personal bill for the other side's fees, still belongs to the lawyer.