A Deadline Extension With a Clear Purpose
On August 12, 2026, the presidency of the Supreme Court of Justice of Buenos Aires Province (SCBA) issued Resolution RP 894/26, pushing back the public comment deadline on its draft "Reglamento para el desarrollo y uso responsable de la Inteligencia Artificial" from August 7 to August 21 (justicia.ar; SCBA). The court cited "diversas solicitudes recibidas tras el cierre del plazo inicial" — requests from institutions that wanted more time to weigh in on a rule the court itself calls institutionally significant. It's a small procedural footnote, but it points to a larger and better-drafted policy than most "AI and the courts" headlines suggest.
The regulation, opened for consultation on June 24 under Resolution SC 1719/26, does two distinct things. First, it declares that jurisdictional decision-making is non-delegable: no judge or official may hand AI the resolution of a case, the weighing of evidence, or the construction of legal reasoning behind a ruling (SCBA). Second, and more novel, it bars uploading personal data, confidential information, or details of ongoing investigations into public or commercial generative AI tools whose terms of service allow that data to be reused for model training — a provision that, as written, functionally rules out free-tier ChatGPT and similar consumer chatbots for anything touching a live case file (Infobae).
The Case for the Bright Line
The strongest argument for the first prohibition isn't hypothetical. Courts in multiple countries have already sanctioned lawyers for filing briefs with AI-fabricated case citations, and evidence-weighing is precisely the kind of context-dependent, high-stakes judgment where a large language model's confident wrongness is most dangerous — a defendant's liberty or a family's custody arrangement can turn on it. Argentina's Data Protection Law 25.326 already obligates confidentiality over personal data; extending that logic to AI vendors whose commercial terms permit training on submitted text is a reasonable, technically literate reading of an existing statute, not a new burden invented from scratch. Judicial systems have a narrower margin for error than most institutions adopting AI, and a rule that keeps humans accountable for "substantive, effective and documented control over results produced" is a proportionate response to a real risk, not a moral panic.
Where Regulators Usually Get This Wrong — and Where SCBA Doesn't
Most judicial AI rules written in the last two years fail in one of two ways: they ban tool use outright and get ignored, or they say nothing and let each judge freelance a policy. A Northwestern University survey of federal judges in the United States, conducted in December 2025, found that of 112 responding judges, roughly 61.6% already use at least one AI tool in their work, yet policy is a patchwork — about 20% formally prohibit it, a third permit or encourage it, and nearly a quarter have no policy at all (Northwestern). That's the predictable result of silence.
The EU has gone the opposite direction: the AI Act classifies any AI system intended to help a judicial authority research facts, interpret law, or apply law to a case as high-risk under Annex III, triggering conformity assessments, technical documentation, and human-oversight obligations that fully bite by December 2, 2027 (artificialintelligenceact.eu). That's a heavier compliance regime than Buenos Aires is proposing — SCBA's draft doesn't require pre-market conformity assessments for every research tool a court might use.
SCBA's draft threads the needle between these poles. It doesn't ban AI-assisted legal research, drafting help, or document summarization — the tools that actually save time. It bans two specific things: outsourcing the decision itself, and leaking case data into tools built to learn from it. That's a scalpel where the EU used a blunter instrument and where most U.S. courts have used nothing at all.
The Real Test Is Enforcement, Not Drafting
The open question — and the reason the extended consultation matters — is what happens after August 21. The draft creates a governance commission to review submissions and report by August 31, but neither the regulation nor the coverage so far specifies penalties for a judge who violates the training-data provision, or an audit mechanism to catch it before a leak happens rather than after. A rule that's unenforceable in practice functions as a suggestion, and suggestions don't stop a overworked clerk from pasting a confidential filing into a free chatbot at 11 p.m.
The consultation process itself is worth crediting on its own terms: SCBA opened comments to bar associations, academic institutions, and the public before finalizing binding text, and extended the window when stakeholders asked for more time rather than closing it on schedule. That's a better model than regulation by judicial memo. If the final version pairs its already-sound distinctions — assistance versus delegation, private versus training-enabled tools — with a concrete audit and reporting mechanism, Buenos Aires will have produced one of the more usable judicial AI frameworks written this year, in any jurisdiction.