The Law Firm That Built Its Own Hallucination Machine

On January 22, 2025, three lawyers filed a motion in limine in federal court in Wyoming, in a case about a hoverboard that caught fire and, by some accounts, burned the plaintiffs' house down. Nine cases anchored the argument. Eight of them were fiction.

That's the whole premise. Let it sit for a second.

The case is Wadsworth v. Walmart Inc., over a Jetson hoverboard Walmart sold that allegedly went up in flames. It landed in front of US District Judge Kelly H. Rankin in the District of Wyoming, which is exactly where you least want a judge discovering that most of your citations are invented.

Here's the twist that makes this one different from the usual AI courtroom disaster. The usual version is a tired associate with ChatGPT open in a browser tab at 2am, quietly hoping nobody would check. Morgan and Morgan built its own AI platform for exactly this purpose, called it MX2. law, and pointed it at Wyoming case law. Most firms get caught using a hallucination machine. Morgan and Morgan got caught having built one.

How the motion got written

Attorney Rudwin Ayala drafted the filing and used MX2. law to generate the case law. His prompts, later disclosed to the court, were about as ordinary as prompts get: add federal case law from Wyoming on motions in limine, then add more. MX2. law obliged. It is described in filings as an in-house database launched by Ayala and Morgan's firm, their own tool, built for their own lawyers. It answered exactly like every other well-formatted and entirely wrong AI output before it. It just did so with a firm letterhead attached.

Nine cases went into the motion. According to one account, eight were invented. According to another, the judge could locate precisely one of the nine. Either way, the math is bad. You do not want your citation success rate to round to zero.

Three attorneys' signatures went on the filing: Ayala, T. Michael Morgan and Taly Goody of the Goody Law Group, serving as local counsel. Morgan and Goody first laid eyes on the motion after it went out. They signed it anyway. That detail is going to matter in about two paragraphs.

The bill

On February 24, 2025, Judge Rankin issued the sanctions order. Ayala had his pro hac vice admission revoked, was removed as counsel of record and was fined $3,000. Morgan and Goody were each fined $1,000 for Rule 11 violations. Total damage: $5,000, plus the professional equivalent of a permanent stain on a good shirt.

The firm itself was spared sanction. Judge Rankin noted Morgan and Morgan had already put remedial steps in place, training, and a requirement that anyone using AI-generated material verify it independently before it goes anywhere near a filing. A rule that, notably, arrived too late to save this particular motion.

The sharpest part belongs to Morgan and Goody, the signers who never wrote a word of it. Signing a document, Judge Rankin wrote, certifies that the attorney read it and made a reasonable inquiry into it. That duty is nondelegable. You cannot hand it to a colleague, and you certainly cannot hand it to a colleague's chatbot. Trusting the associate who trusted the AI is still trusting the AI. It just adds a layer of plausible deniability that a federal judge is free to ignore.

Judge Rankin also reached for a line that should be getting framed and hung in law school hallways at this point: a fake opinion is not existing law, and citing one does not provide a non-frivolous ground for anything, quoting the 2023 Mata v. Avianca sanctions ruling. Two years and one very expensive lesson apart, and lawyers were still finding new ways to arrive at the same sentence.

To its credit, the firm owned much of it. The attorneys withdrew the motion once the order to show cause came down. They were forthcoming about the AI use, they paid opposing counsel's fees and the firm rolled out new training. Ayala told the court this was the first time he had ever used AI this way; which is a strange kind of comfort, the professional equivalent of "I've never crashed a car before" delivered from inside the wreckage. The firm's own statement called the episode "a cautionary tale for our firm and all firms," which is either admirable honesty or the sound of a PR department doing its actual job. Possibly both.

Judge Rankin, for what it's worth, stopped short of writing an obituary for legal AI. He noted that AI can be genuinely useful for attorneys and the public, done right. This was done wrong, at speed, unchecked, by a firm that had built its own confidence machine and forgot to build a way to doubt it.

The scale problem nobody mentions

Morgan and Morgan is a heavyweight. By head count, it ranks 42nd among US law firms. A firm that size filing motions with invented case law is a preview of what happens when a well-resourced firm automates its own research pipeline and skips the step where someone checks the output against reality. MX2. law was a tool the firm built and trusted, at scale, before it had earned that trust.

The rerun

Sanctions have a long memory, it turns out. In May 2026, over a year after the Wyoming order, a Massachusetts state court judge denied T. Michael Morgan's application to appear in a case alleging Harvard University misused human remains. The reason cited was the 2025 Wyoming sanction for AI misuse. Morgan runs the firm's product safety practice. He is licensed in Florida and Kentucky, and he is the son of firm founder John Morgan, all of which weighed nothing, in the end, against one sanctions order from a hoverboard case a year in the past.

That's the part that should genuinely worry anyone who thinks a sanction is a one-time cost you pay and move past. It follows you. It becomes a fact attached to your name, and it travels. A judge in a Massachusetts probate matter, with zero connection to a Wyoming hoverboard fire, read that fact and made a decision because of it. The citation you didn't check costs you more than the case in front of you. It follows you into rooms you haven't walked into yet.

The 2am version

Everyone in this story wanted to serve the court. Ayala asked a tool his firm built for exactly this purpose to find him case law, and it did what these systems do: it produced case-shaped sentences with the right formatting and nothing real underneath. Morgan and Goody signed a document they hadn't read because they trusted the colleague who had. Every individual decision here is one a reasonable, busy person could make on a Tuesday. Stack them and you get eight fictional cases in a federal filing.

This is the trap for anyone who cites a source under deadline, lawyers with in-house AI platforms included. Ask a model for sources on your thesis and it will hand you clean formatting and a plausible author, with a publication year that fits your argument suspiciously well. It reads exactly like every real citation you've ever seen, because it was built from the shape of real ones. You find out it's empty when someone downstream goes looking. Ayala found out from a federal judge. Most people find out from a professor circling a paragraph in red, smaller rooms, same sting.

The fix that would have saved Morgan and Morgan $5,000 and a year of headlines was one boring check: does this case actually exist, and does it say what the filing claims it says. That's the whole gap between a motion that stands and a motion that gets you fined by a federal judge for citing fiction as law. Checking your own citations against the real record, before you file anything, is the one step that turns a hallucination into a footnote instead of a headline.