On September 4, 2026, Argentina's Procurador General Eduardo Casal signed Resolution PGN 64/2026, issuing the first binding, institution-wide rules for artificial intelligence use inside the Ministerio Público Fiscal — the federal prosecutors' office that investigates everything from tax fraud to organized crime. The resolution bars staff from entering case files, sensitive personal data, or information under legal reservation into public-facing AI tools like ChatGPT, Gemini or Copilot; prohibits decisions based solely on automated output; requires mandatory training in prompt formulation, bias detection and result validation; and directs two internal directorates to build secure, in-house AI environments so prosecutors eventually have an alternative to consumer chatbots (Infobae; Microjuris).
The rule didn't come out of nowhere. The office created a dedicated AI Program back in May 2025 under Resolution PGN 14/2025, tasked with drafting technical standards and ethics protocols (Palabras del Derecho). Sixteen months later, that program produced binding guidelines built explicitly on the OAS's Inter-American Data Governance and AI framework, UNESCO's 2021 Recommendation on the Ethics of Artificial Intelligence, the OECD's AI Principles, and NIST's technical definitions (Microjuris).
The Case the Prosecutors Are Making
It's worth stating the strongest version of why this rule exists, because it's a genuinely good one. Federal prosecutors handle grand jury-equivalent material: undisclosed tax audits, ongoing wiretaps, victim identities in trafficking cases. Tax lawyer Julián Ruiz, quoted in the specialist press, warns that uploading ARCA (Argentina's federal tax authority) records to a public chatbot could violate fiscal secrecy protections under Article 157 of the Penal Code — with consequences ranging from procedural nullities that unravel an entire case to breaches of Argentina's information-sharing commitments with the OECD and the IRS (iProfesional). That is not a hypothetical harm. A single prosecutor pasting a case summary into a free consumer tool to get a faster draft could, in principle, blow up years of investigative work and expose the state to liability. Institutions with that much downside exposure are right to move before an incident forces their hand rather than after.
The resolution also solves a coordination problem regulation is genuinely good at solving: without a common rule, AI adoption inside a 4,000-plus-person federal agency was going to be decided office-by-office, prosecutor-by-prosecutor, with no consistency in what data protection looked like from one jurisdiction to the next. A single, mandatory floor — don't feed in case data, always keep a human in the loop, train people properly — is a sensible default for an organization that cannot tolerate the variance.
Where the Caution Should End
Where this deserves scrutiny is in what comes next. The resolution's affirmative promise — that the office will build secure, in-house AI tools "progressively" — is doing a lot of work with no committed timeline. Bans are easy to write and enforce immediately; internal tooling is expensive, slow, and dependent on procurement cycles that Argentina's public sector does not have a strong recent record of moving quickly through. If the secure alternative takes two or three years to materialize, prosecutors are left for that entire window with formally sanctioned AI use that, in practice, has no usable on-ramp — which tends to produce exactly the shadow-IT problem the resolution is trying to prevent, just conducted more quietly.
There's also a broader inconsistency worth naming. The Milei government's pitch to Congress for a national AI law rests on a stated commitment not to regulate the technology prematurely, precisely to keep Argentina attractive as an AI investment and data-center hub (search results, hcdn.gob.ar proyecto de ley 2026). That national bill remains pending in committee. Meanwhile, one federal agency has unilaterally adopted a stricter, risk-averse AI posture than anything currently in statute — grounded in UNESCO's 193-member-state ethics recommendation and the OECD's AI Principles, which 47 countries plus the EU have endorsed (UNESCO; OECD). That's not a contradiction — institutions with confidential data obligations should hold themselves to a higher bar than the general economy — but it does mean Argentina is now regulating AI from the bottom up, agency by agency, while the top-down framework that's supposed to give the whole system coherence sits unresolved. A prosecutor's office is exactly the kind of institution that should move first on data-handling discipline. But if every ministry writes its own AI rulebook while the national law stalls, businesses and citizens will face a patchwork that's harder to navigate than either a single light-touch statute or no statute at all. Casal's office got the internal call right. Congress should treat that as a prompt to finish the external one.