AI hallucinated eight of nine cases in a Morgan & Morgan brief — a federal judge fined the lawyers $5,000 and revoked one pro hac vice

Verification status: PENDING — awaiting human confirmation. Not verified. This is an honest-negative AI-governance case, not a deployment win. Every figure originates with an independent federal court, not a self-reporting vendor.

The problem

Generative AI marketed for legal research can fabricate authority. In Wadsworth v. Walmart Inc. (D. Wyo., No. 2:23-CV-118-KHR), plaintiffs’ attorneys from the national firm Morgan & Morgan filed motions in limine on 22 January 2025 that “cited nine cases, but eight did not exist” (Order on Sanctions, ECF No. 181). On 10 February 2025 the attorneys “admitted the cases were not real and hallucinated by an AI platform.”

What was built

The tool was Morgan & Morgan’s own in-house generative-AI legal-research platform. Per the court, drafter Rudwin Ayala “uploaded the brief onto ‘MX2.law’ to add case law,” a site the order describes as “an in-house database launched by Mr. Ayala and Mr. Morgan’s firm, Morgan & Morgan.” He prompted it to “add … Case law from Wyoming setting forth requirements for motions in limine” and, “[w]ithout verifying their accuracy,” included the fabricated citations. The order records “this was his first time ever using AI in such a way.”

The outcome

On 24 February 2025, U.S. District Judge Kelly H. Rankin found all three attorneys violated Rule 11(b) and imposed sanctions:

• Rudwin Ayala (drafter): “pro hac vice admission is revoked … removed as counsel of record,” plus “a penalty of $3,000 … into the Registry of this Court within fourteen days.”
• T. Michael Morgan (supervising attorney): “a penalty of $1,000 … within fourteen days.”
• Taly Goody (local counsel): “a penalty of $1,000 … within fourteen days.”

The three penalties total $5,000 (the order lists them individually; the sum is not a separate figure in the order). The court declined to sanction the law firms themselves, noting Morgan & Morgan had trained employees not to use the AI software this way and had since added a verification acknowledgement to MX2.law — a disposition both secondaries also carry: Ambrogi wrote that “Judge Rankin declined to sanction the attorneys’ law firm, noting that Morgan & Morgan had trained its employees not to use AI software in the way Ayala had and had since implemented an additional acknowledgment requiring users to independently verify AI-generated information.” This is the same court-adjudicated class as Mata v. Avianca and Coomer v. Lindell: the judge — independent of the sanctioned parties — states the numbers in a signed, publicly filed order. Two newsrooms independent of each other, of the court and of the firms reported the same facts the next day: LawSites (Bob Ambrogi) and Bloomberg Law, whose reporter Bernie Pazanowski wrote that “Ayala had his pro hac vice status in the case revoked and he was fined $3,000. Morgan and Goody were each fined $1,000,” and that the firm itself “wasn’t sanctioned, because it already took steps to insure that its lawyers in the future independently verify any AI-generated information before relying on it.” Ambrogi wrote that “Ayala used his firm’s in-house AI platform, MX2.law, to generate case law when drafting the motions,” and that the three were fined “for their roles in submitting motions in limine that cited eight non-existent cases.” Bloomberg Law does not name the tool — it says only “its AI program” — so MX2.law is corroborated by one newsroom, not two. Neither secondary carries the ECF number, the 22 Jan 2025 filing date, the 10 Feb 2025 admission date or the nine-case denominator; those rest on the order alone.

Path to green

No client-confirmation call is needed to establish the figures: they are stated by an Article III court in a signed order that is byte-tied to Wayback (capture 20260723175444, CDX digest G5WHVISEX7GL44B2BZHN23SG2KOPIKMS) and mirrored on the Internet Archive’s RECAP item. If TIN chooses to carry honest-negatives, the human step is confirming with the Clerk of Court (D. Wyo.) the two things the reachable public record does not settle.

What the docket does show. A CourtListener v4 RECAP capture taken this session (Wayback 20260724020634, CDX digest BLGQHLOVBEA3D6RQYUM63Y5QNCNU3ZHB, byte-tied) records docket 67564066 with "dateTerminated": "2025-07-28": five months after the sanctions order the case ended by “STIPULATION of Dismissal WITH PREJUDICE … pursuant to Fed.R.Civ.P. 41(a)(1)(A)(ii)” (ECF No. 246), granted the same day by “the Chief District Judge Kelly H Rankin” (ECF No. 247). No notice of appeal appears in the capture.

What it does not show, and why the human step remains. First, Registry payment is a Clerk-of-Court fact rather than a docket fact — no entry recording payment of the three penalties appears in the evidence reachable here, and the docket may never record one; the v4 search endpoint returns only top-matching entries per docket and the docket-detail endpoint is auth-walled (HTTP 401), so this is an open question, not a negative finding. Second, no appeal on this docket is not no appeal — a Rule 11 sanction appeal can sever into a separate appellate or mc docket. A Tenth Circuit RECAP search for this case name returns no matching docket, but RECAP’s appellate coverage is contributor-dependent, so that is a weak negative and not proof. A PACER/Clerk check settles both.