AI Hallucination Cases

Defendant's reply brief contained citations that did not support her arguments and included at least one non-existent case citation; the court concluded these errors suggest use of AI and treated the issues as abandonment under Rule 28(b)(6), dismissing the appeal.

The court found the plaintiff's filings contained misrepresentations of the record—specifically, a purported 'Clerk's Entry of Default' that never existed and a claim that an Immigration Judge made a final binding APA finding. The court attributed these misrepresentations to likely AI generation, warned the plaintiff about consequences for false statements, and construed the misrepresentations as AI misapplication rather than deliberate deception.

The appellant (self-represented) cited two nonexistent cases in her appellate brief. The respondent flagged the fictitious citations and requested sanctions. The court found the citations fictitious, discussed sanction authority and AI-generated filings, but declined to impose sanctions because the request was procedurally inappropriate, the appellant corrected filings promptly, and the legal propositions were, in fact, supported by existing authority.

Order to show cause is here here.

"In total, appellant's opening brief contains 23 case quotations, 21 of which are fabrications. Appellant's reply brief contains many more fabricated quotations. And, both briefs are peppered with inaccurate citations that do not support the propositions for which they are cited.

[...]

We conclude by noting that "hallucination" is a particularly apt word to describe the darker consequences of AI. AI hallucinates facts and law to an attorney, who takes them as real and repeats them to a court. This court detected (and rejected) these particular hallucinations. But there are many instances-hopefully not in a judicial setting-where hallucinations are circulated, believed, and become "fact" and "law" in some minds. We all must guard against those instances."

The appellate court observed that the pro se appellant's opening brief cited cases that do not exist and exhibited hallmarks of generation by a large language model (repetitive 'refined' drafts, internal suggestions, and the statement 'Generative AI is experimental'). The court identified the fabricated citations and noted the brief's deficiencies but proceeded to decide the jurisdictional timeliness issue on the merits, affirming dismissal.

The appellant admitted using generative AI to prepare his Notice of Appeal and Summary of Argument. The Court found several cited authorities could not be located (concluding they were fictitious) and held that deploying unverified AI-generated research that cites non-existent cases breaches duties not to mislead the court and risks contravening Pt XIVB of the Family Law Act 1975 (Cth) (on confidentiality of proceedings).

The court found nearly every citation in counsel's motion to be incomplete, inaccurate, or fabricated, describing the references as typical of AI hallucinations and ordering counsel to show cause why sanctions should not be imposed.

The court found multiple exhibits (videos, photographs, messaging screenshots, and metadata) to be fabricated or materially altered using generative AI. The court deemed Plaintiffs' explanations not credible, declined criminal referral, declined monetary sanctions, and imposed a terminating sanction under Cal. Civ. Proc. Code § 128.7(b).

The court granted Plaintiff's motion to strike portions of the Opening Brief after Defendant raised concerns that the brief included a non-existent quotation attributed to an existing case, a mischaracterization of an existing case, a citation to a non-existent case, and a miscitation of a case that did not address the asserted issue. The court noted counsel had been sanctioned in a separate case for citation-related deficiencies consistent with AI-generated hallucinations. The stricken portions were removed and the ALJ decision was affirmed.

"In her declaration, Attorney Cervantes described the erroneous citation as her own inadvertent error and further implied that Westlaw’s tool may have “glitched” by erroneously producing the citation. [Dkt. 31 at ¶ 4]. Specifically, she indicates that the copy citation tool mayhave inadvertently copied the wrong citation information.

Yet, that explanation seems highly improbable. The citation was incorrect in every respect and none of the sub-parts of the citation match each other, including the case docket number, reporter, pincite, court information, and date information. The likelihood of every component of a citation being simultaneously wrong due to a software malfunction in transcribing from a “correct” citation is, in the Court’s view, statistically improbable (particularly given Westlaw’s representations as to the accuracy of its tools and the apparent resources Westlaw has devoted to promoting the reputation of this tool). If one set of numbers had been transposed, or if the date were wrong, such transcription errors might be explicable. However, the fact that the party names, docket number, Westlaw citation, date, and court all failed to match, resulting in a mashup citation, cannot credibly be attributed to a software error, let alone a mistranscription or copying mistake. Rather,this Frankensteinian legal citation, stitched together from mismatched party names, docket numbers, dates, and courts, bears the hallmarks of an AI-generated hallucination, as documented in numerous published opinions and reports."

El tribunal verificó que varias sentencias invocadas por la recurrente no pudieron ser halladas en repertorios oficiales; asumió, por el beneficio de la duda, que ello provino del mal uso de herramientas de IA y advirtió la obligación profesional de verificar fuentes. Se impone recomendación al letrado y se declaran inoficiosos sus trabajos en el recurso (sin condena disciplinaria formal).

See the story here.

The magistrate judge found multiple cited authorities in the plaintiff's complaint to be fabricated or misquoted: two cases could not be located and two quotations attributed to real cases did not appear in those opinions. The court warned the plaintiff and ordered him to provide copies of cited cases in future filings; recommended dismissal of federal claims and possible sanctions for continued misrepresentations.

Plaintiff admitted using AI to prepare filings, submitted an unsigned affidavit for a purported attorney and cited cases the court could not locate; court found false statements and potential fake caselaw and ordered show-cause re: Rule 11 sanctions and evidentiary hearing.

Father admitted using ChatGPT to generate case citations; 14 cited cases were non-existent or misattributed. Court treated them as AI hallucinations, disregarded them, awarded costs, and ordered declarations for future AI use.

Counsel's Amended Motion cited a non-existent case (In re Berry Good, LLC) generated by AI. Court found the citation fabricated, counsel admitted AI use and negligence, implemented safeguards, and court declined sanctions while transmitting matter to disciplinary office. The court's reflections are worth quoting in full.

"When examined, paragraph 15 is the result of a flawed AI search that suggested an incorrect generalization falsely supported by a specific but non-existent case. Under any scenario, a suspect argument buttressed with a case from whole cloth violates the expressed and tacit norms governing the practice of law as it intersects with the courts. Simply, you cannot make stuff up to convince a court to do something.

In simpler times, that would end the inquiry with draconian consequences. The times, however, are no longer simple; thus, the inquiry does not end here.

AI interposes additional considerations that defy easy categorization or scrutiny when examining the traditional norms of practice and advocacy. Threshold, two fields of inquiry present in any AI debate. The first is whether exposition artificially generated by a computer is acceptable in what has contextually been a forum for human critical analysis, thought, and advocacy. That is not the instance here. Rather, it is the second; that is, when AI sacrifices accuracy to satisfy the consumer and that inaccuracy is advocated before the court.

[...]

The only acceptable conclusion is that citing a made-up case is a false representation to the court that violates the expressed norms of practice, is actionable, and should bear consequences. The answer, however, to the second question —whether an artificial component alters the calculus—is more difficult.

Here, our traditional norms are challenged. In context, the traditional advocacy process begins with black letter law in the form of statutes, codes, and regulations. Then, there are judicial opinions and orders interpreting black letter law in the context of justiciable issues based on discrete facts. Lawyers cite cases for their binding or persuasive authority and have done so ever since someone started committing judicial decisions to stone, parchment, or paper. The ability to find and cite cases has evolved over time from laborious research in courthouses, to compendiums, indices, digests, reporters, advance sheets, key numbers, and now the internet.

The internet changed this dynamic in two ways. First, it streamlined, refined, and improved the research process. Users could select fields, topics, and key numbers to find and locate cases. Eventually, more specific tools, such as key words or phrases, enhanced the process. Second, and more pertinent to the present issue, AI is creating a process where instead of just finding authority, the computer also does the supplicant's thinking and analysis for them. The first innovation continues to evolve; the second is new, likewise continues to evolve, but constitutes an entirely new dimension, rather than a continuum, in legal practice. Finding has evolved into finding and thinking.

Travel back to research, the first changed dynamic. From stone to the internet, how lawyers found and used case law took many forms. Some would carefully read every case they found; some would read every page of select cases they thought pertinent; others might read the digest entry and then read only that section of the case; others might read only the digest entry; others barely read anything at all and cited anything that looked like it might support their position. Each was a personal decision by the lawyer for which he or she is accountable under the traditional norms of practice both officially—through codes of conduct and Rule 9011—and unofficially per the norms and expectations of courts.

AI compresses this historical dynamic. The research engine is now the search and thinking engine. AI finds case law and tells the subscriber that it supports their position. All the personalities described above—and known oh-so-well to us all—have become one.

Except, however, AI is flawed. In its infancy—albeit one that may be in college soon—it is immature and does not always provide accurate information or correct analysis; it may even make things up.

A tool that potentially supplants your advocacy by doing the finding and thinking for you is enticing. 3 But in the end, it is only a tool. Despite our near total and unfortunate faith in the internet, it does not relieve the attorney of his or her responsibility to make sure that the information and analysis are correct.

Reliance on AI can, however, mitigate intent. Here, Counsel is responsible for their negligence and misplaced reliance on AI, but there is no indication—and this court does not believe —that they purposely misled the court. They did not decide to make up a case. Misplaced reliance or negligence does not always equal intentional misrepresentation."

In her latter response to the Court, Counsel apologised for her use of a "a non-firm approved general AI service".

The Special Master found that Dr. Sawyer cited non-existent sources produced by an AI tool without adequate verification. The court concluded the opinions themselves were otherwise reliable and denied exclusion, while allowing cross-examination and noting that an award of costs might be appropriate.