Puerto Rico Supreme Court Sets AI Hallucination Standard for Lawyers

Signing a Brief with Hallucinated Case Law Is No Longer a Correctable Error. Puerto Rico's Supreme Court Has Made It a Canon Violation.

Six case citations. Five non-existent or misattributed to the Supreme Court of Puerto Rico (TSPR). A motion to dismiss filed on July 17, 2025. And behind it all, a generative AI tool that produced text with the full appearance of legal authority — and none of the substance. In re María V. Irizarry Centeno and Anissa M. Bonilla Irizarry, AB-2025-0232, decided on April 21, 2026, is the first case of its kind in Puerto Rico and one of the most analytically complete in the Spanish-speaking world: the TSPR not only sanctions the conduct — with a formal warning (apercibimiento) rather than suspension, given significant mitigating factors — but uses the moment of normative transition to articulate a doctrine on technological competence and diligence that will outlast the specific facts by a wide margin.

The ruling's importance does not rest on the severity of the sanction. It rests on the doctrinal architecture built by both the majority Opinion of Associate Justice Candelario López and the concurring Opinion of Associate Justice Rivera Pérez, joined by Associate Justice Estrella Martínez. Together, these two texts offer the first Puerto Rican judicial doctrine on what technological diligence means, when AI use constitutes sanctionable professional conduct, and against what parameters future sanctions should be calibrated. What follows is a doctrinal analysis of both opinions.

The Citations That Never Existed: Anatomy of the Failure

Before examining the doctrine, it is necessary to understand the factual texture of the case — because the morphology of the error is precisely what determines its legal characterization.

On July 17, 2025, attorneys Irizarry Centeno and Bonilla Irizarry filed a motion to dismiss on behalf of a co-defendant in a labor case before the San Juan Court of First Instance. The brief contained six citations presented as express pronouncements of the TSPR, complete with case names, reporter volumes, page numbers, and quoted text in quotation marks. None were authentic: some referred to cases that do not exist under that name or citation; others cited real cases whose referenced pages contained no such text; still others attributed propositions to decisions that addressed entirely unrelated matters.

Opposing counsel flagged the issue on August 4, 2025, filing a motion denouncing the fictitious case law and requesting sanctions for recklessness. The following day — and this speed would prove a decisive mitigating factor — the attorneys filed a Clarifying Motion acknowledging that some citations had been "erroneously attributed due to an involuntary error in the research and selection of case law," attributing the problem to "limitations arising from consulting general internet sources." On August 8 they filed an Informational Motion substituting the erroneous citations with authentic authority. On August 22 the trial court issued the order that triggered this disciplinary proceeding: it referred the matter to the TSPR, imposed a $1,000 sanction under Rule 9.3 of the Civil Procedure Rules for causing unjustified delays, and directed the Supreme Court to evaluate the attorneys' professional conduct.

Justice Rivera Pérez devotes specific attention in her concurrence to the typology of the error — and this distinction is legally decisive. What occurred here was not a debatable legal interpretation, an overextended analogy, or an improbable but arguable theory. What occurred was the textual attribution, in quotation marks, with volume and page numbers, of pronouncements that the TSPR never made. In the Justice's formulation, the filing introduced into the adjudicative process "hallucinated authority produced by algorithmic fabrication" — text with the appearance of binding precedent but devoid of any normative force, because its "normative force" as case law presupposes a real justiciable controversy adjudicated through a real adversarial process.

Canon 18: The Duty of Competence Absorbs Technology

The canonical architecture of the ruling rests on three pillars: Canons 18, 35, and 38 of Puerto Rico's Code of Professional Ethics. The first — competence and diligence — carries the most doctrinal weight because it is the one reformulated in technological terms.

Canon 18 requires lawyers to render "suitable, competent, careful, and diligent" representation. Prior TSPR case law had defined the competence duty as referring to the obligation to "possess the legal knowledge necessary for the adequate handling of a case," linked to staying current in one's knowledge of the law. The 2026 ruling does not abandon that definition — it extends it: a lawyer who is not current on the technological tools shaping contemporary legal practice, including generative AI, also fails the competence duty.

The most significant element is the formulation of the negative standard: the duty of competence "is not satisfied by the mere intention to argue in accordance with the law, but requires a reasonable and diligent effort to corroborate the accuracy and authenticity of the authorities cited." This sentence operates a subtle but crucial transformation: it converts source verification into a constitutive element of the Canon 18 duty, not an optional best practice. If you do not verify, you do not comply. Full stop.

Justice Rivera Pérez develops this point with greater technical depth in her concurrence, citing In re Jackson Hospital & Clinic, Inc. (Bankr. M.D. Ala. 2025): the output of a large language model depends heavily on the prompt, which in turn requires the attorney to have a foundational understanding of the legal issue at hand. A prompt based on an incorrect assumption about the law — or biased toward a particular result — may steer the attorney not only away from actual legal authority, but also away from any plausible legal argument supporting their position. Technological competence is therefore not about knowing how to operate a tool; it is about interrogating the tool from a solid legal knowledge base and verifying its responses with critical independence.

Here is where the analysis becomes most interesting: technological competence and technological diligence are conceptually distinct, though functionally inseparable. The first is static — knowing the capabilities, limitations, and risks of the tools. The second is dynamic: using those tools with timeliness, responsibility, and continuous supervision in each specific engagement. It is possible to have competence without diligence — a lawyer who fully understands how an LLM functions and still does not verify its outputs — and even, though less commonly, diligence without sufficient competence. Canon 18 requires both.

Canon 35 and Objective Truth at the Moment of Signing

Canon 35 establishes the duty of sincerity and honesty before courts, prohibiting the use of means "inconsistent with the truth" and inducing the adjudicator into error "through artifice or a false account of facts or law." Prior TSPR precedent had established that this canon "is breached by the mere fact of an attorney departing from the truth, regardless of the reasons therefor," and that proving deliberate intent or the intent to defraud is unnecessary.

Applying this standard to the present case yields a consequence that cannot be softened: the subsequent correction of the citations, the transparency with which the attorneys acknowledged the error, the good faith with which they claimed to have acted — none of this erases the Canon 35 violation already consummated when the brief was filed. The duty of sincerity and honesty "requires that representations made to the court be accurate at the time of filing." Candor, under this canon, is not retroactive.

This has systemic consequences the TSPR does not fully explore but which deserve emphasis. If a Canon 35 violation requires no mens rea and is consummated the moment inaccurate information is filed, then any use of generative AI without prior verification that produces non-existent citations constitutes, per se, a potential canonical violation. The tool may be sophisticated; the hallucination may be convincing; the good faith may be total. None of this matters under Canon 35 if the text presented to the court was inaccurate. AI is not an excuse — not even a mitigating factor — for the sincerity duty. It operates, with some margin, only as a factor in calibrating the sanction.

Canon 38: The Appearance of Propriety as a Collective Good

Canon 38 requires lawyers to uphold the honor and dignity of the legal profession, avoid the appearance of improper conduct, and identifies as a violation any conduct that "affects the moral conditions of the attorney and makes them unworthy of belonging to this forum." The TSPR applies this canon with relative restraint: the filing of multiple non-existent or misattributed citations "projects an image of professional carelessness incompatible with the values of honor, dignity, and responsibility demanded of the legal profession." And, in a phrase worth underlining: "the trust that society places in the institution of justice diminishes when an attorney's conduct represents impropriety."

What is at stake under Canon 38 is not only the individual attorney's reputation — it is the institutional credibility of the adversarial system itself. A court that receives false case law must dedicate resources to verifying what should have been verified before filing; opposing counsel must combat authorities that do not exist; the adjudicative process is distorted. Generative AI used without adequate supervision is not just a risk to the client: it is a negative externality that affects the administration of justice as a whole.

Rule 9.1 and the Solemnity of the Signature

The majority opinion gives explicit attention to Rule 9.1 of the Civil Procedure Rules, which imposes on the signatory of a filing the "affirmative duty to conduct a reasonable investigation of the applicable facts and law before presenting the brief." The provision operates at the procedural level — its breach may trigger sanctions under Rule 9.3 — but the TSPR integrates it into the canonical analysis as an element that gives substance to the duties of competence and diligence.

The connection matters: signing a legal filing is not a formality — it is a substantive certification. The signatory certifies that the brief "is well-grounded in fact and supported by existing law." If the law cited does not exist, the certification is false, regardless of the origin of the error. AI cannot sign: only the person who assumes responsibility for what is signed can. And that responsibility must be earned through the verification Rule 9.1 demands.

Rule 1.19 and the New Standard of Technological Competence

One of the ruling's most significant elements from a prospective standpoint is the explicit invocation of Rule 1.19 of Puerto Rico's new Rules of Professional Conduct, approved June 17, 2025 and in force since January 1, 2026. The rule was not applicable to the facts — the TSPR is careful to note that the conduct is governed by the 1970 Code of Professional Ethics — but it is invoked as an interpretive canon and as evidence that the doctrine being constructed is not ad hoc: it is the jurisprudential expression of principles already codified.

Rule 1.19 expressly introduces the duties of "technological competence and diligence," defining competence as the duty to "acquire the necessary skills and maintain reasonable knowledge of technological developments that may impact the practice of law." The official commentary defines technological diligence as the obligation to "use technological tools or solutions in a timely, responsible manner without causing undue delays or prejudice," and explicitly warns that "presenting documents without validating content generated through technological tools may constitute a lack of diligence, even when one possesses technological competence."

This bifurcation — competence without diligence as an autonomous violation — is Rule 1.19's most precise normative contribution to the problem the case presents. Knowing what an LLM is and understanding how it works is not enough: every output must be verified in each specific matter. Technological competence is a necessary but not sufficient condition; diligent verification in practice is the specific act the legal order demands — and the one the attorneys omitted.

Justice Rivera Pérez adds a dimension the rule's text does not explicitly capture: the diligence duty in reviewing AI-generated content "does not end upon filing; it continues when the attorney reaffirms or defends the positions contained therein." If the attorney subsequently discovers that a citation is false, non-existent, or incorrect, the duty to correct without delay is immediate. The attorneys in this case acted correctly on this point — they filed the Clarifying Motion the very next day after being notified — and this contributed materially to mitigating the sanction.

Comparative Doctrine: From Mata v. Avianca to the Mattox Framework

Justice Rivera Pérez's concurrence is, in substantial part, an exercise in applied comparative law: it reviews the line of federal cases on generative AI misuse — with Mata v. Avianca, 678 F.Supp. 3d 443 (S.D.N.Y. 2023) as the foundational precedent — and extracts from it criteria proposed for future Puerto Rican disciplinary decisions.

Mata v. Avianca established that filing ChatGPT-generated fictitious decisions constitutes an "abuse of the adversary system" sanctionable under Federal Rule of Civil Procedure 11 — analogous to Puerto Rico's Rule 9.1 — for two reasons: the attorneys failed to conduct a reasonable inquiry before filing and presented as existing law decisions that no court had ever issued. The sanction was $5,000, with additional corrective measures. The operative standard is categorical: "A fake opinion is not 'existing law' and citation to a fake opinion does not provide a non-frivolous ground for extending, modifying, or reversing existing law."

The concurrence integrates the more elaborate framework of Mattox v. Prod. Innovations Rsch., LLC, 807 F. Supp. 3d 1341 (E.D. Okla. 2025), which offers three structural factors for evaluating the conduct: (1) verification and inquiry — whether counsel conducted a reasonable, human-based verification of every cited authority before filing, and whether reliance on the automated tool alone satisfies that standard (it does not); (2) candor and correction — whether counsel disclosed AI use and corrected the record promptly once inaccuracies were discovered; and (3) accountability and supervision — whether supervising attorneys exercised adequate oversight and whether institutional safeguards existed to prevent recurrence.

Justice Rivera Pérez proposes five parameters specific to Puerto Rican jurisdiction, functioning as sanction-graduation criteria: (1) materiality of the citation — whether it supports a central proposition or resolves a substantive controversy; (2) appearance of precision and degree of falsity — whether it was presented with quotation marks and a specific volume-and-page reference, versus a reasonable paraphrase of existing doctrine; (3) pattern or repetition — whether there is a single isolated citation or multiple false entries in the same or other filings; (4) prior reasonable investigation — whether direct human review occurred; and (5) level of experience and technological knowledge of the attorney. Though formulated in a concurrence — and thus without binding ratio decidendi force — these criteria constitute the most articulated interpretive guide available for future cases.

Why the Warning Was the Right Sanction Here — And Will Not Be Next Time

The TSPR applied the mitigating factors with notable precision. The attorneys had no prior disciplinary record. They filed the Clarifying Motion the day after being notified. They paid the procedural sanction without delay. They withdrew from representing the client. They expressed sincere remorse. The case was, functionally, an isolated incident inconsistent with their professional trajectory. And crucially, the TSPR acknowledges it faces this typology of conduct for the first time: the absence of specific precedent in the jurisdiction — combined with the post-facto entry into force of Rule 1.19 — justifies an initial warning sanction rather than suspension.

But the ruling itself, with a clarity that leaves no interpretive margin, warns that "conduct of this nature will not be tolerated" going forward and that "repetition of such conduct could result in severe disciplinary sanctions," including suspension of practice. The warning is not a signal of permissiveness — it is proportionality calibrated to a first offense in a normative environment that was still consolidating. The same threshold of tolerance, applied today — with Rule 1.19 in force, the ruling published, and the doctrine fixed — would in all likelihood produce a materially different outcome.

The Problem the Ruling Does Not Resolve: Plausible Hallucinations and Genuine Arguments

The analysis would be incomplete without identifying the limits of the doctrine developed. The ruling describes with clarity a scenario of maximum hallucination opacity: citations presented with every appearance of precision — case name, volume, page, quotation marks, verbatim text — that are wholly fabricated. But the hallucination phenomenon is more varied and sometimes harder to detect.

What happens when the cited case exists but the quoted language is a model-generated elaboration that does not appear in the actual decision? Or when the doctrine attributed to a case is reasonably consistent with that court's jurisprudential line, though not expressly stated in the cited decision? Or when the hallucination is not one of citation but of normative interpretation — the model "invents" the content or scope of a statutory provision in ways that pass a superficial verification check? The concurrence hints at these nuances in distinguishing between "textual citation with quotation marks and a specific volume-and-page reference" and "a reasonable paraphrase of existing doctrine," without fully developing the legal consequences of each scenario.

That said, the logic of the ruling is extensible: the verification duty requires not only confirming that the volume and page number exist, but that the quoted text — or the attributed proposition — actually corresponds to what the decision says. The signature certifies substance, not form. And in that extension lies perhaps the most demanding task the new standard imposes on the profession: not just finding the real citations, but reading them.

Five Practical Implications for Law Firms

The April 2026 ruling functions as a trigger for transformation in the working habits of any firm using generative AI for legal research and drafting. The implications are at least five.

First: source verification is now a canonical duty, not a best-practice recommendation. Citing case law without independent verification in a primary database — Westlaw, LexisNexis, the TSPR's own portal, LEXJURIS — carries real disciplinary risk.

Second: the prompt used to interrogate a generative AI tool is part of the exercise of legal competence, not a delegable technical matter. A poorly framed, biased, or legally incorrect prompt does not just produce false citations — it can steer the attorney away from any plausible legal argument.

Third: supervisory responsibility in multi-attorney firms — including supervising partners overseeing associates who use AI — is squarely implicated. The Mattox accountability-and-supervision factor applies directly to law firm organizational models. Absence of an institutional AI policy does not immunize; it implicates.

Fourth: prompt correction and transparent disclosure, once an error is discovered, carry significant mitigatory weight. The attorneys' behavior in this case — Clarifying Motion filed the next day, sanction paid without delay — was decisive in grading the sanction. Delay or denial, as the cited federal case law shows, operates in precisely the opposite direction.

Fifth: although the ruling is specific to Puerto Rican jurisdiction, its argumentative structure — verification, signature as certification, Canon 35 without mens rea, sanction graduation — is directly transferable to any Latin American legal system with analogous professional responsibility rules. Justice Rivera Pérez's concurrence, with its systematic survey of federal comparative law, provides a roadmap that bar associations and supreme courts across the region would do well to examine.

The AI Does Not Sign, But Whoever Signs Is Responsible

The essential doctrinal contribution of In re Irizarry Centeno and Bonilla Irizarry can be stated in one sentence: generative AI is a support tool that does not modify, and cannot modify, the structure of professional responsibility of the attorney. It can expand capabilities, accelerate research, suggest argumentative frameworks. But it cannot substitute professional judgment, and it cannot sign.

Whoever signs, certifies. And that certification requires, under Canons 18, 35, and 38 and Rule 9.1 of the Civil Procedure Rules, that the law cited exist, that the citations be authentic, and that the prior investigation have been reasonably diligent. It does not matter that the error was unintentional. It does not matter that the tool generated text of impeccable appearance. It does not matter that the underlying doctrine was, in general terms, correct. What matters is that the text presented to the court was inaccurate in the specific, and whoever signed it assumed responsibility for that inaccuracy.

The TSPR has drawn the line. What remains to be seen is how quickly the profession — and the bar associations that regulate it — internalize the new standard. Rule 1.19 has been in force since January 1, 2026. The first disciplinary ruling is published. The warning is unambiguous. The margin for claiming ignorance has closed.

Case analyzed: In re María V. Irizarry Centeno and Anissa M. Bonilla Irizarry, AB-2025-0232, 2026 TSPR 41 (April 21, 2026). Author: Associate Justice Candelario López. Concurring opinion: Associate Justice Rivera Pérez (joined by Associate Justice Estrella Martínez).

Download full ruling (PDF)