Skip to content
ContestabilityA research
publication

Research note / 2026-11-18

A signature is not a review. What the new Ibero-American AI standards count as human intervention

What was approved

On 12 November 2026 the Spanish data protection agency announced that the Ibero-American Data Protection Network, whose permanent secretariat it holds, had unanimously approved standards on the use of personal data in AI systems. The agency calls the text the first international instrument to treat the subject in a full articulated form. The text is in Spanish, and the English below is my paraphrase. By its own terms (section 1.1.2) it is orienting guidance that complements national law and does not replace it.

Three provisions I read against the NIC questions

The first concerns notice. Saying that AI or algorithms are used does not satisfy the duty to inform (2.5.1). The information has to let a person understand the data categories and sources, whether the system produces profiles, inferences or decisions, the degree of human intervention, and the routes for exercising rights and challenging the outcome (2.5.2). Where it is needed, a meaningful explanation covers the logic, the factors that weigh materially, the data sources, the known limits, and how errors and bias are detected (2.5.3). Source code is not required, and trade-secret limits cannot be used to take away the understanding a person needs.

The second concerns review. A decision stops counting as exclusively automated only if the human intervention is prior, real and effective, by someone able to understand the output and with authority to confirm it, change it or depart from it (4.2.1). A reviewer who only formalises or endorses the output does not meet that test, and alert tools cannot stand in for the reviewer (4.2.2). The notice, for its part, may not present a merely formal supervision as a safeguard (2.5.2).

The third concerns inferences. Where an inference cannot be corrected directly, the person may ask for correction of the data behind it, add information, contest it or ask for review (5.3).

Why this is a contestability question

The standards place the test on the reviewer’s capacity, not on the presence of a person. That is where I have been placing it, and Paper A lists sociotechnical roles and evidence control among its six candidate functions. The connection is under development; I do not claim the Standards adopt the NIC framework.

A proposed illustration

This is a hypothetical. A benefits applicant is told that a caseworker reviewed her claim. Suppose the caseworker had ninety seconds per file and no authority to depart from the score. Under 4.2.1 the decision would remain automated, and under 2.5.2 the notice could not honestly call the review a safeguard. Were the claim decided in the EU, Article 22 of the GDPR would also be in play. Whichever national law applied would govern the case, not the Standards.

What the source does not say

Section 7.9 makes the documentation available to the supervisory authority. In the sections I read, I did not find a duty to preserve the system version and inputs as they stood on the day of the decision, or a way for the person to see the reviewer’s record. I have not read every provision. Nor does the AEPD note report any application of the text.

If a notice must now say what the reviewer could do, who will check that the sentence is true?