On May 22, 2026, Argentina's Ministry of Human Capital unveiled Gemelo Digital Social — a "Social Digital Twin" meant to model, anticipate and forecast the impact of social policies before they are rolled out. Minister Sandra Pettovello posted that "Argentina se adelanta al futuro" (Argentina gets ahead of the future); President Javier Milei declared that "for the first time, our country leads the social future." The system is designed to fuse employment, education, childhood, social-program, demographic, provincial and economic records — plus data from private actors — into a single model that runs along four axes: descriptive, explanatory, predictive and prescriptive. The government calls the destination a "predictive State."
Within 48 hours, national deputy Agustín Rossi (Unión por la Patria, Santa Fe) filed a formal information request, and digital-rights specialists warned that a system of this scope had been announced with no published governance framework, no legal basis and no data-protection assessment. That gap — not the ambition — is the problem.
The case for forecasting is real
It is worth stating the government's strongest argument plainly, because it is a good one. Social spending in Argentina has long been reactive, fragmented across overlapping registries, and vulnerable to both leakage and exclusion errors. A well-built model that simulates how a transfer, a training program or an eligibility change will land before it is implemented could reduce waste, catch unintended harms early, and target scarce resources toward the people who need them most. "Digital twins" already do exactly this for energy grids and transport networks. Using evidence to design social policy is not dystopian; it is what competent governments should aspire to. A publication that is pro-innovation should not flinch at the State using AI well.
The objection, then, is not to the tool. It is to the sequence.
Governance is supposed to come first, not in "phase four"
The Ministry's own rollout plan is revealing. According to La Nación, questions of "privacy, algorithmic ethics, data governance and applicable legal frameworks" are slated for a fourth development phase — after the architecture, the data integration and the partnerships are already built. That ordering inverts how data-protection-by-design is supposed to work. Once millions of records have been fused into a single model, retrofitting consent, purpose limitation and access controls is far harder than designing them in.
This matters more in Argentina than the announcement acknowledged. The country's personal-data statute, Ley 25.326, dates to 2000 and predates modern machine learning entirely. The data-protection authority (the AAIP) has spent years preparing a reform aligned with the GDPR — one that would introduce accountability-by-design and, crucially, a right to object to automated decisions that produce legal effects. That bill lost parliamentary status and remains stalled. So the Social Digital Twin is being assembled on top of a legal framework that has no clear rule for exactly the kind of large-scale, cross-database automated profiling the system performs.
What Rossi is actually asking
Rossi's filing (expediente 0138-D-2026) is not a demand to halt the project. It poses roughly fifteen questions: the legal and administrative acts authorising the system, the contracts and tenders, the interoperability protocols, the budget committed so far, the international actors involved, any privacy-impact assessments, the cybersecurity and external-audit arrangements, and whether the AAIP issued an opinion. He asks pointedly whether the system will reuse data collected for other purposes without fresh consent, and whether it will automate decisions about who can access social programs.
Those are the right questions, and the fact that they can be asked at all — and that the answers are apparently not yet public — is the tell. Purpose limitation (data gathered to administer a pension should not silently become a fraud-risk signal) and a ban on opaque automated eligibility decisions are not anti-innovation hurdles. They are the conditions under which a forecasting tool stays a forecasting tool rather than drifting into individual-level surveillance.
The Netherlands already ran this experiment
Argentina does not have to speculate about how this goes wrong. On February 5, 2020, the Hague District Court struck down the Dutch SyRI system, which had cross-referenced employment, tax, benefits and housing data to flag citizens for fraud investigation. The court found it violated Article 8 of the European Convention on Human Rights because it was insufficiently transparent and lacked adequate safeguards — and investigators noted it had been deployed mainly in low-income neighbourhoods, concentrating suspicion on the poor. SyRI was a narrower system than a full social digital twin, and it was still ruled unlawful.
The lesson is not "don't build it." It is that opacity plus scale plus the most vulnerable population equals a system courts and the public will eventually reject — usually after the reputational and human damage is done.
A proportionate path forward
None of this requires Argentina to abandon the predictive-State ambition. It requires three commitments that any pro-innovation government should welcome, because they make the tool more durable, not less:
- Publish the legal basis and a data-protection impact assessment before integrating live citizen data, and have the AAIP review it — the governance that is currently "phase four" should be "phase zero."
- Codify purpose limitation and a prohibition on fully automated benefit denials, with a human in the loop and a right to explanation and appeal.
- Keep the model at the population level for forecasting, with strong barriers against re-identification, rather than scoring named individuals.
Milei is right that the future does not wait. But a predictive State that cannot show its citizens the rules it runs on is not ahead of the future — it is one court ruling, or one leak, behind it. Argentina has a rare chance to build this the right way round. The cheapest moment to do so is now, before the data is fused.