Wadsworth v. Walmart: Why an In-House AI Tool Is Not a Defense

A federal judge in Wyoming revoked one lawyer's pro hac vice admission and fined three attorneys after they filed a motion citing nine cases, eight of which did not exist. The fabricated authority came not from a consumer chatbot but from the firm's own in-house artificial intelligence platform. That single fact is the reason this case deserves a careful read inside any firm that is buying, building, or evaluating legal AI.

What happened in Wadsworth v. Walmart

The underlying matter was an ordinary products-liability suit. The plaintiffs, Stephanie and Matthew Wadsworth, alleged that a hoverboard sold through Walmart and manufactured by Jetson Electric Bikes caught fire and burned their home. The case is Wadsworth v. Walmart Inc., No. 2:23-CV-118-KHR, in the United States District Court for the District of Wyoming.

On January 22, 2025, plaintiffs' counsel filed a set of motions in limine. The motions cited nine cases. When opposing counsel and the court looked for them, eight of the nine could not be found, because they had never been decided. They were fabricated citations, complete with plausible reporter numbers and parentheticals, the hallmark output of a generative model asked to supply supporting authority.

The drafting attorney, Rudwin Ayala, had used his firm's in-house platform, identified in the record as MX2.law, to add case law to the brief. He uploaded the draft, queried the tool for supporting authority, and incorporated what it returned. He did not pull the cases, did not read them, and did not confirm they existed before the motions were filed.

On February 24, 2025, United States District Judge Kelly H. Rankin issued the sanctions order. The court found that the attorneys had violated their certification obligations under Federal Rule of Civil Procedure 11(b), which provides that a lawyer who signs and files a paper certifies, after a reasonable inquiry, that the legal contentions are warranted by existing law. Citing authority that does not exist is the clearest possible failure of that certification.

The sanctions were tiered to reflect each lawyer's role:

• Rudwin Ayala, who drafted the motions and used the tool, had his pro hac vice admission revoked and was ordered to pay a $3,000 fine.
• T. Michael Morgan, the supervising attorney from Morgan & Morgan whose name was on the filing, was fined $1,000.
• Taly Goody, local counsel, was fined $1,000.

To their credit, the lawyers did not dig in. The court noted that the respondents were forthcoming, honest, and apologetic, and that they had taken remedial steps, including withdrawing the offending motions, before the court ruled. That candor almost certainly held the sanctions to the modest level imposed. It did not erase the revoked admission or the public order that now carries the firm's name.

The detail that matters: it was the firm's own tool

Coverage of the Morgan & Morgan AI case has tended to lead with the firm's size and the embarrassment of a prominent plaintiffs' shop being caught with invented citations. The more instructive detail is quieter. The fabricated cases did not come from a free, general-purpose chatbot used on a personal account. They came from MX2.law, an internal platform operated by the firm itself.

This matters because the most common reaction to the wave of AI-hallucination sanctions has been a reassuring story that goes something like this: those lawyers used the wrong tool. They typed into a consumer chatbot instead of a real legal product, and a purpose-built legal platform would have protected them. Wadsworth is the case that breaks that story. A platform built or branded for legal work, used inside a large and sophisticated firm, still produced eight nonexistent cases that landed in a federal filing.

The lesson is not that in-house or legal-specific tools are bad. It is that the tool was never the safeguard. The safeguard is a verification step performed by a human before the document is filed, and that step was missing regardless of which software generated the text. Empirical work supports the broader point. A 2024 Stanford RegLab and Human-Centered AI study, "Hallucination-Free? Assessing the Reliability of Leading AI Legal Research Tools," tested purpose-built legal research products and found that they still produced incorrect or unsupported information at meaningful rates, on the order of roughly 17 percent for one major product and higher for another. Retrieval and legal training reduce hallucination. They do not eliminate it. A capable tool changes the odds. It does not change the duty.

This was not the first warning, and it will not be the last

Wadsworth sits in a now-established line of cases. The first widely reported example was Mata v. Avianca, Inc., 678 F. Supp. 3d 443 (S.D.N.Y. 2023), in which Judge P. Kevin Castel sanctioned two attorneys and their firm $5,000 after they submitted a brief containing citations invented by ChatGPT and then, when questioned, failed to come clean promptly. Courts across multiple jurisdictions have since issued comparable orders, and the pattern of facts is remarkably consistent: a lawyer asks a model for authority, the model obliges with realistic-looking citations, and no one checks before filing.

The professional-responsibility framework was also in place well before the Wyoming order. On July 29, 2024, the ABA Standing Committee on Ethics and Professional Responsibility issued Formal Opinion 512, "Generative Artificial Intelligence Tools," the association's first formal ethics guidance dedicated to generative AI. The opinion does not tell lawyers to avoid these tools. It maps the existing duties onto them: competence under Model Rule 1.1, including a reasonable understanding of a tool's capabilities and limitations; candor toward the tribunal under Rule 3.3; confidentiality under Rule 1.6; supervision of subordinate lawyers and nonlawyer assistance under Rules 5.1 and 5.3; and reasonable fees under Rule 1.5. Opinion 512 is explicit that the output of a generative tool must be evaluated and verified by the lawyer. Wadsworth is what that paragraph looks like when it is ignored.

For a firm of three to thirty attorneys, the takeaway is not that you face exotic new rules. It is that familiar rules now have a new and very efficient way to be violated. Rule 11 has always required reasonable inquiry. The change is that a junior associate, or a partner under deadline, can now generate a fully formatted brief with supporting citations in seconds, and every one of those citations may be fictional.

Build the protocol the tool cannot provide

The defense Wadsworth forecloses is "we used a real legal AI tool." The defense it leaves open is "we have a verification protocol, we followed it, and we can show our work." Small firms are well positioned to put that protocol in place precisely because they are small. You do not need a committee. You need a short written policy and the discipline to follow it.

A workable verification protocol includes the following:

• Independent confirmation of every citation. Every case, statute, and quotation in a filing must be confirmed in a primary source or an established reporter database before the document goes out. The standard is not "the tool said so." The standard is that a human opened the authority.
• Read the case, not just the citation. A citation can be real while the proposition it supposedly supports is invented. Confirm that the authority exists and that it actually says what the brief claims.
• A named signer who owns the inquiry. Under Rule 11, the lawyer who signs certifies the filing. Assign a specific person to perform and attest to the citation check before any signature, and make clear that supervising attorneys, as in Wadsworth, are accountable for work filed under their names.
• Disclosure of AI use inside the team. Associates and staff should flag when generative tools contributed to a draft, so the reviewer knows where to look hardest. This is a workflow norm, not an admission of wrongdoing.
• Confidentiality screening before input. Opinion 512 cautions against putting client confidences into tools that may use or expose that data. Know how any platform, in-house or vendor, handles what you type into it, and obtain client-informed consent where the tool's data handling requires it.
• A standing policy that survives turnover. Put the rules in writing so they apply to the next lateral hire and the next contract attorney, not just the people who happened to read the latest sanctions order.

None of this requires you to forgo AI. Used with a verification layer, these tools genuinely accelerate drafting, summarizing, and first-pass research. The firms that get in trouble are not the ones using AI. They are the ones treating the tool's confidence as a substitute for their own inquiry.

The bottom line for managing partners

The Wadsworth v. Walmart AI sanctions are not a story about one careless lawyer or one flawed product. They are a story about a category error. The error is believing that the right software discharges the duty to verify. It does not, and a court has now said so in an order that revoked an admission and fined a partner whose only act was signing the brief.

Treat any AI platform, including one your own firm builds or licenses, as a drafting assistant whose output is unverified until a lawyer confirms it. Write that expectation down, assign it to a person, and audit it. The cost of the protocol is a few minutes per filing. The cost of skipping it, as Wadsworth shows, is measured in fines, lost admissions, and a published opinion with your firm's name in the caption.

This is general information for lawyers and law-firm leaders, not legal advice, and it does not create an attorney-client relationship. The authorities are cited so you can read them yourself.