By This Hour Business Technology Desk
A lawyer who used ChatGPT was reportedly punished after citing testimony attributed to witnesses who did not exist, an allegation that goes beyond the now-familiar concern that generative AI can produce false legal authorities. The reported problem was not merely an inaccurate summary or an unhelpful draft. It involved purported testimony from invented people being presented in a legal setting.
The account, reported by Ars Technica, places the central failure at the point where AI-generated text moved from a drafting tool into material relied on by a lawyer. If accurately described, the episode shows how a system capable of producing fluent, confident prose can create a particularly acute risk when its output resembles the record of a real proceeding. A fabricated witness does not simply weaken an argument; it can alter the factual basis on which an argument appears to rest.
Only a limited account of the matter has been supplied for this report. It identifies a lawyer, use of ChatGPT, punishment, fabricated testimony and made-up witnesses, but it does not provide the lawyer’s name, the court, the jurisdiction, the type of proceeding, the wording of the sanction, or the sequence by which the false material was discovered. Those omissions materially limit what can be concluded about the case and the severity of the conduct alleged.
The reported error concerns facts, not just legal research
Public discussion of AI use in law has often focused on invented case citations: text that looks like legal authority but cannot be found. The report at issue describes a different alleged category of failure. The false material was said to be testimony, and the supposed witnesses were said to have been made up. That distinction matters because testimony is presented as an account from a person, while a witness is presented as a real individual connected to the factual record.
For a lawyer, references to testimony can perform several functions at once. They can be used to characterize events, support a requested conclusion, challenge another account, or give a filing the appearance of close engagement with a record. When the people attributed with the statements do not exist, none of those functions can safely be assumed. The issue is not only whether a sentence is true or false. It is whether the asserted source for that sentence is real.
The reported punishment indicates that the matter was treated as more than an internal drafting mistake. Still, the available claim does not say what the punishment consisted of. It would be unwarranted to infer a fine, a contempt finding, a professional-disciplinary outcome, a change in the underlying case, or any particular finding about intent. Nor does the supplied material establish whether ChatGPT independently generated the false content, whether a user prompted it in a way that contributed to the result, or whether text was changed after it was produced.
Those distinctions can be significant in assessing responsibility, but they cannot be resolved from the limited record available here. What the report does establish, if its description is accurate, is the claimed connection between ChatGPT use and citations to fabricated testimony from nonexistent witnesses, followed by punishment of the lawyer involved.
A verification failure can survive polished language
Generative AI’s usefulness and its risk arise from the same feature: it can rapidly produce text in a form that appears complete. A draft that names people, attributes statements and follows the tone of a legal filing may look more finished than it is. Apparent specificity can be especially misleading when the underlying details have not been checked against the relevant materials.
The reported case therefore turns on verification rather than on whether a chatbot can assist with writing in the abstract. A system may be used to organize prose, identify questions or propose language, yet none of those uses removes the need for a lawyer to establish that factual assertions and cited sources are genuine. Where a filing identifies testimony and witnesses, the risk is heightened because the text purports to describe people and evidence rather than simply offer a line of reasoning.
That is also why a broad label such as “AI error” would obscure the reported conduct. The report attributes use of ChatGPT to the lawyer, but the lawyer was the person reportedly punished. On the information provided, there is no basis to characterize the chatbot as a party to the proceeding or to suggest that the tool, rather than its user, bore the legal obligation at issue. The episode is better understood as an alleged failure to validate material before relying on it.
For businesses developing or deploying generative-AI tools, the account highlights a practical boundary that product fluency cannot solve by itself. A tool may produce text that is persuasive in form while lacking a reliable connection to the record a user is expected to know. Warnings, instructions and review processes may affect how users approach that problem, but the supplied reporting does not describe what safeguards, if any, were available or used in this instance.
The account leaves the disciplinary record largely unknown
The word “punished” carries clear significance but little precision without an order or fuller description. It could refer to a judicial response, a professional consequence, or another form of sanction. The supplied claim does not identify the decision-maker, the legal standard applied, the amount of any monetary penalty, or whether the lawyer was given an opportunity to explain the filing. It also does not say whether the reported action is final.
Likewise, the number of fabricated witnesses and the amount of supposed testimony are not available in the material provided. A single invented attribution and a filing built around multiple fictitious people would pose different questions of scale, though both would be serious if presented as genuine. The absence of those details means the report cannot reliably describe the breadth of the alleged fabrication.
Nor is there enough information to establish the effect on any client or on the underlying dispute. The available account does not say whether a court relied on the false material, whether an opposing party identified it, whether a filing was withdrawn or corrected, or whether the matter affected an outcome. Such questions are central to understanding practical harm, but they remain unanswered here.
A separate report previously published by this outlet concerned a New Mexico defense lawyer who was reportedly fined after an appeal filing was found to contain invented witnesses and police testimony. That account may be useful context because it describes the same broad problem of fabricated people and evidence entering court papers, but the limited material supplied here does not establish that it concerns the same lawyer, proceeding or punishment. Readers can review that earlier report on the New Mexico filing for its separately stated details and limitations.
Why the distinction matters for AI use in legal work
The reported episode is a warning about the difference between assistance and substantiation. In legal work, a draft can suggest an argument, but claims about people, testimony and records require an evidentiary foundation outside the generated text. Treating a chatbot’s specificity as proof risks importing its unsupported assertions into a process that depends on identifiable sources.
The case also illustrates why errors involving invented factual material may be harder to catch than vague or obviously implausible prose. A name, a witness label and a purported statement can look internally coherent. The text may even fit the argument a user hopes to make. Yet coherence is not confirmation. The relevant question is whether the named person and the attributed testimony can be located in the actual record.
That is a limited conclusion, but it is the one supported by the report. The available information does not show that every use of ChatGPT in legal practice produces false material, or that the tool cannot be used responsibly. It instead describes a reported instance in which its use was associated with an alleged failure serious enough to bring punishment after fabricated testimony was cited.
Further documentation could clarify the case considerably. A court order, disciplinary decision or filing could establish the lawyer’s identity, the nature of the proceeding, the exact text at issue, the stated reason for the punishment and any explanation offered in response. It could also show whether the tribunal made findings about how ChatGPT was used, rather than simply noting its presence in the preparation of the material.
Until such material is available, the account should be read narrowly. Ars Technica reported that a ChatGPT-using lawyer was punished for citing fabricated testimony attributed to made-up witnesses. That report has not been independently corroborated by this publication, and the absence of accessible source-page context prevents independent assessment of the underlying documents, procedural history and exact sanction.
Reporting notes
What is confirmed: The reported allegation links ChatGPT use, fabricated testimony, made-up witnesses and punishment of a lawyer.
Why this matters: The report raises concerns about verification when AI-assisted legal writing asserts facts, sources or people as real.
What remains unclear: The lawyer, court, jurisdiction, sanction, scope of the false material and procedural outcome were not provided. This report is based on one source and has not been independently corroborated.
