AI Hallucination Cases

"[T]he fact that Plaintiffs … did not "fabricate cases or cite nonexistent decisions" is of no help. When a case cite is "real," an attorney, or for that matter a judge, might see a case they recognize and assume the quote or holding has been accurately represented. That problem is illustrated here; although Chat GPT generated "holdings" that looked like they could plausibly have appeared in the cited cases, in fact it overstated their holdings to a significant degree. And while a litigant might get away with similar overstatements because they could, perhaps, reason their way to showing how a case's stated holding might extend to novel situations, an LLM does not reason in the way a litigant must.

To put it in a slightly different way, LLMs do not perform the metacognitive processes that are necessary to comply with Rule 11. LLMs are tools that "emulate the communicative function of language, not the separate and distinct cognitive process of thinking and reasoning." When an LLM overstates a holding of a case, it is not because it made a mistake when logically working through how that case might represent a "nonfrivolous argument for extending, modifying, or reversing existing law or for establishing new law;" it is just piecing together a plausible-looking sentence—one whose content may or may not be true."

Court later declined to reconsider this decision (see here).

"We disagree with respondent, however, that dismissing the appeal is an appropriate sanction for Farivar’s conduct. Our inherent authority to impose this sanction “should be exercised only in extreme situations, such as where the conduct was clear and deliberate and no lesser sanction would remedy the situation.” (Crawford v. JPMorgan Chase Bank, N.A. (2015) 242 Cal.App.4th 1265, 1271.) We conclude that we can sufficiently address the prejudice to the parties and the court from [Counsel]’s sanctionable conduct and sufficiently achieve the deterrent purpose of sanctions (see, e.g., Code Civ. Proc., § 128.7, subds. (b)(2), (c) & (h)) by doing the following:

First, [Counsel] shall pay sanctions in the amount of $7,500 to the clerk of this court within 30 days after the remittitur is filed. We calculate this amount based on, inter alia: (1) the significant amount of time this court spent verifying the fabricated citations in the opening brief, and (2) that Farivar refused to accept responsibility for his conduct, instead characterizing the fabricated quotations and citations as mere “clerical citation errors” and continuing to misrepresent legal authority in his opposition to the sanctions motion.

Second, we strike appellant’s opening brief and require appellant to file, within 10 days of the issuance of this order,a corrected opening brief. Appellant’s corrected brief may differ from the version originally filed only to the extent it corrects or omits the fabricated citations and quotations in the original version. Appellant shall file and serve both a final version of the new brief as well as a redline version.

Finally, because we conclude attorney Farivar has violated a Rule of Professional Conduct, we are required to “take appropriate corrective action.” (Cal. Code Jud. Ethics,canon 3D(2).) In line with this obligation, we direct the clerk of the court to serve a copy of this order on the State Bar.

We acknowledge and have considered that, as appellant argues, the majority of the fabricated quotes in the opening brief do not appear to be misrepresentations that work to appellant’s advantage; that is, the brief does not represent the law to be more favorable to appellant’s arguments than it actually is.

Nonetheless, we must consider broader concerns about the integrity of the courts and the legal profession. Inaccurate citations in briefing—whether the result of technological hallucinations or human failure to verify—may be relied on in court decisions, “circulated, believed, and become ‘fact’ and ‘law’ in some minds. We all must guard against those instances. . . . ‘There is no room in our court system for the submission of fake,10hallucinated case citations, facts, or law. . . . ’ [Citation.]” (Noland, supra, 114 Cal.App.5th at pp. 448-449.)"

Plaintiff repeatedly cited non-existent or misleading authorities in her Fourth Amended Complaint. The Court identified at least one fabricated citation and several mischaracterized cases, noting these failures to comply with Rule 8 and prior court instruction, and relied on the deficient pleadings in granting dismissal.

Show Cause Order is here.

See here for details.

Order to Show Cause is here.

"No puede soslayarse, como hecho de gravedad institucional, que la defensa técnica del imputado - en el momento procesal que constituye la última instancia de revisión extraordinaria local- haya presentado un escrito elaborado posiblemente mediante el uso de herramientas de inteligencia artificial, u obtenidos a través de medios no confiables y fidedignos sin ejercer el más mínimo control sobre la veracidad de las citas jurisprudenciales, la autenticidad de las fuentes invocadas ni la coherencia dogmática de los fundamentos empleados.

La conducta descripta trasciende el mero error material o descuido profesional: configura un supuesto de negligencia grave incompatible con los deberes de probidad, lealtad y diligencia que rigen la funcióndel abogado defensor.

La defensa penal no puede convertirse en unespacio de experimentación de medios y/o tecnologías de manera irresponsable que - aunque de utilidad potencial - requieren un manejo prudente, crítico y siempre supervisado por el criterio humano. El ejercicio de la defensa en juicio, protegido por el art. 18 de la Constitución Nacional y por los instrumentos internacionales con jerarquía constitucional, exige una intervención personal, reflexiva y fundada de quien ejerce la representación técnica. El uso automático y acrítico de un asistente artificial, o la utilización de fuentes con procedencia desconocida para producir un escrito en la instancia recursiva vulnera no solo el deber de diligencia, sino también el derecho del imputado a una defensa real y efectiva."

"[134] Même si l’arbitre de griefs est un tribunal administratif moins formel que le sont laCour supérieure ou les autres tribunaux judiciaires, notamment parce que ses règles depreuve sont plus souples, il n’en demeure pas moins que c’est un tribunal. À l’évidence,le procureur qui, devant ce tribunal, s’appuie sur de la jurisprudence doit s’assurer qu’elleexiste.

[135] L’arbitre de griefs s’attend à ce que tous les procureurs qui plaident devant luisoient compétents, honnêtes, professionnels et respectueux de son autorité.Manifestement, celui qui soumet au tribunal des références jurisprudentielles inexistantesne satisfait pas ces attentes, car il induit, intentionnellement ou non, le tribunal et la partieadverse en erreur.

[136] Qui plus est, le procureur qui fait référence à de la jurisprudence qui n’existe pasrallonge inutilement l’arbitrage. On reproche déjà, souvent avec raison, la longueur etles coûts élevés associés à l’arbitrage de griefs. Ces problèmes seront exacerbés si lesinformations inexactes générées par les hallucinations d’outils d’intelligence artificielles’introduisent devant les tribunaux d’arbitrage en raison de la négligence des procureurs.Le présent cas en est un bon exemple.

[137] Enfin, le procureur qui fait référence à de la jurisprudence inexistante expose lapartie qu’il représente à devoir compenser les dommages que cela pourrait causer àl’autre partie.

[138] En définitive, référer à des décisions qui n’existent pas, comme l’a fait laprocureure patronale dans le présent dossier, est un geste répréhensible qui ne devraitjamais se produire en arbitrage de griefs. Ce comportement est d’autant plus grave quecette procureure est membre de l’Ordre des conseillers en ressources humaines agréés."

Plaintiff's counsel filed an opposition brief containing several quotations and citations the Court could not verify, including at least one the Court described as a "fabricated quotation." Defendants first identified the potentially inaccurate citations in their reply. The Court ordered disclosure about AI use, production of drafts, metadata and timesheets, and set a hearing to determine responsibility and whether sanctions or bar referral are warranted.

Show Cause Order here. Law firm explained what happened here.

In the ultimate order, the court noted that "In terms of competence, the threat to attorneys using generative artificial intelligence platforms powered by large language models is two-fold. First, danger exists that the attorney does not understand how the technology functions, believing that the output is real instead of “realistic-looking."."

In finding that the law firm acted with integrity, the Court noted with approval that it had repaid the other side's fees, to the tune of 55,721.2 USD.