The Oklahoma Bar Journal December 2024

employee incurred moving expenses in reliance on the offer of employ ment?” In a lengthy response, the generated text stated without author ity , “A prospective at-will employee who is terminated prior to begin ning work may have a claim … for promissory estoppel.” The response then cited several estoppel cases that had nothing to do with at-will employment, prospective employ ees or promissory estoppel in the employment context. Most concern ing, however, is that the response then cited to what it represented as a court opinion recognizing a prospec tive employee’s reasonable reliance; however, the citation was not to a court’s opinion but to a plaintiff’s petition that was voluntarily dis missed. So instead of responding that the prompt called for an answer that would be a matter of first impression in Oklahoma, Lexis+ AI generated an answer that was part “ungrounded,” part “misgrounded” and part legally incorrect. As attorneys, we recognize that both correctness and grounded ness are extremely important in our work. Unfortunately, when researchers from Stanford eval uated generative AI tools geared specifically toward legal research, the tools failed in one or the other dimension fairly frequently. The researchers found that Lexis+ AI provided both correct and grounded responses on only 65% of queries, and Thomson Reuters’ Ask Practical Law AI did so only 18% of the time. 11 In sum, if you use legal generative AI tools, it is just as imperative to look out for fake cases as it is to check for irrel evant cases – or a lack of support altogether – and you must also check the propositions themselves for accuracy.

Ultimately, some use of generative AI may be inevitable, making it that much more important to understand how it works and what its limitations and risks are.

article identify two dimensions of legal hallucinations: correctness and groundedness. 8 Information provided by an AI model may be factually correct and relevant to the question/prompt, but the informa tion may be “misgrounded” – in that the cited sources are mis interpreted or inapplicable – or “ungrounded” – where no sup porting citations are provided to support the response. 9 The follow ing is an example of a correct but “misgrounded” proposition: The right to same-sex mar riage is protected under the U.S. Constitution. Miranda v. Arizona , 384 U.S. 436 (1966). 10 The sentence is factually correct, but the citation should be Obergefell v. Hodges , not Miranda v. Arizona . A slightly different problem can occur when the prompt asks a legal question that has yet to be answered in the jurisdiction spec ified in the prompt. For example, the authors posed the following prompt to Lexis+ AI: “Does a prospective at-will employee who is terminated prior to beginning work have a claim against the employer for promis sory estoppel when the prospective

writing. The type of hallucination most readers may be familiar with is when AI tools provide citations to nonexistent cases. This occurred in the highly publicized New York case Mata v. Avianca, Inc. , where an attorney used ChatGPT while drafting a court filing that ended up containing citations to multiple fake cases. 4 When ordered to pro duce the opinions themselves, the attorney turned again to ChatGPT, which produced fake opinions. One such fake opinion, “ Varghese v. China Southern Airlines Co., Ltd. , 925 F.3d 1339 (11th Cir. 2019),” bore a docket number and Federal Reporter citation associated with completely different (real) cases, and it even listed real judges as its purported authors. 5 This fake opinion, submitted to the court by the hapless attorney, included internal citations and quotations from yet more fake cases. 6 Some citations in the fake opinion had correct names and citations but did “not contain the language quoted or support the propositions for which they [were] offered.” 7 Importantly, AI tools can hallu cinate in other ways besides pro viding completely fake citations. Authors of a forthcoming research

Statements or opinions expressed in the Oklahoma Bar Journal are those of the authors and do not necessarily reflect those of the Oklahoma Bar Association, its officers, Board of Governors, Board of Editors or staff.

DECEMBER 2024 | 9

THE OKLAHOMA BAR JOURNAL