Estudio Garrido

By Mario Covarrubias Jurado — Estudio Garrido Abogados

In a legal world shaped by the uncritical use of AI, California has enacted a basic rule into law: whoever signs is accountable. SB 574 (Senate Bill 574) requires lawyers to verify AI-generated citations and to disclose their use of AI to the courts, and it prohibits arbitrators from delegating their decisions. It also shows that confidentiality is not merely an ethical duty: it can become a material barrier to accessing secure tools. The Quebec precedent, Argentine case law and recent Chinese guidelines show that the problem is no longer technological, but professional.

California decided to regulate a practice that already pervades the profession: the use of generative artificial intelligence in briefs, research and decisions. Against a backdrop of increasingly frequent sanctions for non-existent citations—and of an arbitral award annulled because of fabricated sources—the new law does not merely require the responsible use of AI: it also requires acknowledging when AI was used and under what conditions confidentiality was preserved.

This intersection of jurisdictions matters because it shifts the discussion from technological convenience to professional responsibility. The question is no longer whether lawyers and arbitrators may use AI, but under what conditions its use remains compatible with the signature, with confidentiality, with adversarial debate and with the personal duty to decide.

A Quebec arbitrator based his award on three judgments that do not exist, an article that could not be located and an award that was never rendered. He stated that he had personally drafted the reasoning and had asked an AI tool to add citations, which he did not verify afterwards. On April 22, 2026, the Superior Court of Quebec annulled the award: the arbitrator had delegated his decision-making function.

According to Mayer Brown, it was the first publicly reported case of an award being annulled because of AI-generated reasoning. The problem knows no borders: what a Canadian court sanctioned in April, a
California statute expressly prohibited in September.

That statute is SB 574 (Senate Bill 574), signed into law on September 30, 2026 by Governor Gavin Newsom: one of the first U.S. laws to expressly regulate the use of generative AI by lawyers. The legislature passed it without a single dissenting vote, and it will take effect on January 1, 2027.

For those of us who have followed Argentine case law over the last fourteen months, the law works as external confirmation: the standard that the Rosario Court of Appeals, the National Court of Appeals in Federal Administrative Litigation and the Superior Court of Justice of Chubut had been building case by case was not a local eccentricity. We had already warned that the uncritical use of AI erodes the reasoning of young lawyers and exposes a profession running on autopilot.

One idea runs through both the California statute and the Argentine decisions: the signature cannot be delegated. The lawyer is accountable for the brief he signs; the arbitrator, for the decision he renders. He need not personally draft every word or renounce AI, but he cannot transfer to a tool the responsibility for professional judgment. And the law adds a thought-provoking corollary: what cannot be delegated should not be concealed either.

What Changes with SB 574?

The law prohibits lawyers from delegating the practice of law to an AI system and requires them to verify its outputs and correct anything erroneous. No court filing may contain citations that the responsible lawyer has not personally verified, even if a machine supplied them. A signature was already mandatory—section 128.7 of the Code of Civil Procedure required that at least one attorney sign in his or her name—; what is new is what the signature now vouches for. Using AI as an aid is not prohibited: what is prohibited is turning the signature into a mere formality.

In addition, lawyers must disclose to the court the use of generative AI in the documents they file, and they may not enter confidential, personally identifying or other non-public information into systems whose access is not restricted to the lawyer and to persons the lawyer authorizes under confidentiality obligations. In practice, the rule forces firms to organize their use of AI: to choose reliable tools, limit access and define which information may not be uploaded. Confidentiality ceases to be solely a professional obligation and comes to depend also on concrete organizational decisions.

Saying That It Was Used

Of all the obligations under SB 574, the duty to disclose is the one that says the most about our times. Generative AI has become entrenched in law firms, but acknowledging it is far less common than using it: an open secret, as if admitting it were confessing an inadequacy. That silence comes at a price: what is not disclosed is not supervised or verified either, and that is where the errors we already know of come from. The law reverses the presumption and turns tacit use into disclosed use.

It is not a confession of guilt. A lawyer who reports having used a tool, and who answers for its output, is not discredited: he submits to the scrutiny that every professional act should be able to withstand. The difficulty lies in the scope of the duty. The text does not specify how disclosure must be made or what kind of use triggers the obligation, and it defines generative AI in broad terms; consequently, indiscriminate disclosure could stop being informative and become noise. Argentina currently has no equivalent obligation, although one could be grounded in the principles of good faith and procedural fairness. It deserves a frank discussion.

Arbitrators: The Most Serious Warning

The most novel part of SB 574 does not reach lawyers alone: it also targets arbitrators, and there the warning is more serious. Arbitrators may not delegate any part of their decision-making process to a generative AI tool, nor rely on AI-generated information outside the record without first disclosing it to the parties and, to the extent possible, allowing them to comment. Arbitration rests on a promise: that the decision will be made by a person chosen for his or her judgment and experience. An award drafted in substance by a language model belies that promise.

Quebec shows the problem in its purest form. The arbitrator dismissed, on a preliminary basis, a claim of approximately CAD 1.2 million for having been filed outside the 90-day contractual time limit, and the sources supporting that conclusion were either non-existent or misattributed: three court decisions that do not exist, a scholarly reference that could not be located, a fictitious award and a real decision that did not say what the award claimed it said. The court did not declare AI incompatible with the arbitral function, nor did it hold that every erroneous citation justifies annulling an award; it held something more precise: AI may assist, but it may not replace the reasoning. According to Mayer Brown, although the arbitrator claimed to have drafted the award himself and to have used the tool only to add citations, without validating them, the court concluded, on the facts proven in that case, that he had delegated his function.

In Esquel, a judge inadvertently left a trace of the tool in his judgment—"here is point IV re-edited, without citations and ready to copy and paste"—and the Esquel Criminal Court of Appeals annulled the conviction on its own motion. The Superior Court of Justice of Chubut reversed that annulment, but was categorical: final responsibility lies with whoever signs. That is the value of the precedent: not the annulment, which did not stand, but the discussion about who is accountable.

Three situations should be distinguished: using AI for instrumental tasks, using it to analyze evidence or legal issues, or delegating decision-making judgment to it. Only the third lies at the core of the concern. And the fact that a model generated a draft does not, by itself, prove that the arbitrator delegated his decision: what is decisive is whether he exercised his own judgment, verified the grounds and assumed responsibility for the award.

The question for Argentine arbitral institutions is an uncomfortable one: what do their rules currently say about the use of AI by arbitral tribunals? If the answer is "nothing," it would be wise to start writing it before a dispute writes it for them.

Argentina: What It Has Already Said and What Is Missing

In a previous paper we argued that using AI without verification currently constitutes a breach of the duty of diligence, and we proposed minimum protocols for verification, traceability, transparency and institutional guidelines. The Rosario court stated that delegating legal research and incorporating its results without cross-checking them is "risky and even reckless"; in Medicus, the Court of Appeals urged the lawyer to verify his citations and referred the matter to the Ethics Tribunal; Chubut declared that AI is a support tool, never an autonomous source of decision, and that whoever signs is accountable.

These precedents converge with California, but they do not amount to a regime like California's: there, there are specific and uniform obligations; here, scattered decisions of uneven scope, almost all concerning lawyers and one concerning a judge. There are grounds for demanding professional accountability for the use of AI; what is missing is a specific framework that defines its scope.

China: When a Supreme Court Issues Guidelines on AI

The judicial response to artificial intelligence is not limited to California or to cases in which its use produced procedural errors. China offers a different institutional perspective. On September 7, 2026, the Supreme People's Court published the Opinions on the Trial of AI-Related Dispute Cases According to Law, a 24-provision document specifically intended to guide the judicial resolution of disputes involving artificial intelligence.

Its relevance lies not only in the matters it addresses—civil liability, intellectual property, personal data and personality rights, among others—but also in the decision to establish specific judicial criteria for a technology that is already generating concrete disputes. According to the institutional presentation of the document, it is the first instrument of its kind issued by a country's highest judicial body.

The Chinese initiative and the California statute respond to different questions. California establishes express obligations for lawyers and arbitrators in the performance of their functions; China provides guidelines for resolving AI-related disputes. They are not equivalent instruments, nor do they pursue exactly the same objective, but both show that the law can no longer treat these problems as an exclusively technological matter.

For Argentina, the comparison is useful for an additional reason. While our courts are beginning to outline, case by case, the duties of diligence and responsibility associated with the professional use of AI, the Chinese experience invites us to ask whether it would be advisable to move toward more systematic judicial criteria. The point is not to import solutions from other legal systems, but to identify which problems require consistent and predictable answers.

Two Limits of the Law

It does not mandate use. SB 574 does not state that a lawyer who forgoes AI is negligent. That discussion remains open, and it is the one we have raised by arguing that failing to use reasonably available tools may also constitute a negligent omission.

It has a hidden cost: compliance also requires infrastructure. The confidentiality requirement may push toward tools with access controls and contractual guarantees that not everyone can afford. This is not an inevitable consequence of SB 574, but a risk of its practical implementation: that the small firm, the public defender or the colleague practicing in the provinces may be placed at a disadvantage compared with organizations that have greater resources. Steve Leben warned, in a broader debate on confidentiality and AI, that poorly calibrated security requirements can turn a reasonableness rule into an access barrier for those who practice with fewer resources. The question also deserves to be raised here. One possible answer is institutional: bar associations, public prosecutors' and defenders' offices and the judiciary could procure licenses with confidentiality guarantees and make them available.

What Can Your Firm Do This Week?

That inequality does not eliminate the duty; it requires managing it with simple rules and reliable tools. None of this requires waiting for a law.

The first duty is the oldest one: whoever signs verifies, citation by citation, and whatever cannot be confirmed in the official source is not cited. Next comes organizing use, with a short list of authorized tools and a clear prohibition on uploading data of clients, parties or witnesses to tools that do not guarantee confidentiality. And every significant filing must have an identified lawyer who answers for its content, regardless of who—or what—drafted it.

What remains is what gives the whole its consistency: a minimal log recording which tool was used, for which task and who reviewed the output—if a filing is challenged tomorrow, that log is the defense—and a disclosure policy that defines in advance when and how the use of AI is reported to the client and, where appropriate, to the court. Resolving this calmly is preferable to improvising it in front of a judge.

The Signature as the Unit of Responsibility
From Rosario to Sacramento, from Esquel to Quebec, from China to Sullivan & Cromwell—which acknowledged AI-generated errors in a filing before the judge—and to Medicus, the cases tell the same story: AI has already entered legal practice, but it has not displaced the profession's most elementary duty. It can help search, organize, compare or draft; it cannot turn the signature into a formality or verification into blind trust.

California needed a statute to say so. China chose to formulate judicial guidelines. Argentina can still decide whether it needs a specific rule or more systematic institutional criteria, or whether it should begin by accepting—in law firms, courts and bar associations—that the use of AI must be explainable. The challenge will not only be verifying citations, but creating real conditions for using AI without sacrificing confidentiality or equality. What is decisive, with or without a statute, is the same: the tool may assist; the signature cannot be delegated.