Theory of Change

This document describes the causal logic behind the work of iaS (Inteligencia Artificial Segura), that is, how and why our activities produce the change we seek; it complements the Vision, which explains what iaS is and where it is heading, and it builds on the activities detailed in What we do.

1. The problem

Spanish is spoken by more than 650 million people, about 520 million of them as a native language, spread across the five continents (Instituto Cervantes, 2025); yet the decisions about how the most advanced AI is developed, deployed and controlled are made mostly in English. We identify four gaps that limit AI safety in the Spanish-speaking world:

1. An evidence gap. There are no measurements of our own on how AI systems behave in Spanish; the available multilingual studies show safety inconsistencies across languages and across risk categories (the M-ALERT benchmark evaluates five languages, Spanish among them, with 75,000 prompts), and the comparison table (Leaderboard) from ODESIA measures an average effectiveness gap of 20% (±4%) between Spanish and English on language processing tasks, according to our review on September 19, 2026. Not even the industry’s most ambitious language initiative treats safety as a dimension of the language problem: Google, which in September 2026 stated that it already covered 300 languages and 86% of the world’s population, concentrates all of that effort on capability and accessibility (speech recognition, translation, offline devices), without once mentioning safety or alignment evaluation as part of that work (Google, 2026); without a hub that documents these failures in Spanish and takes the evidence to regulators and developers, it is hard for anyone to identify and fix them.
2. A training gap. The reference training in AI safety, including that of BlueDot Impact, is delivered in English; for those who work in policy, engineering, law or journalism in Spanish, the path from curiosity to a professional contribution runs through a language barrier that has nothing to do with their ability.
3. A public entry gap. The demand exists, since around one hundred people registered per session at iaS sessions from May to July 2026, but it depends on someone offering, consistently and in Spanish, a place to learn and discuss without needing prior technical knowledge.
4. An institutional gap. When an administration, a university, a company or a media outlet in the Spanish-speaking world wants to understand the risks of advanced AI, it needs a counterpart that speaks its language and knows its regulatory framework; iaS aims to be that counterpart, and to that end it has taken part in the European Commission’s public consultation on the classification of high-risk systems under the AI Act.

2. Our theory

iaS holds that lasting change in AI safety in the Spanish-speaking world comes from three elements that we consider insufficient in Spanish, namely, evidence of our own, people who understand the risks and know how to act, and a stable organization that institutions and media can talk to; the most effective way to build them is a common infrastructure that operates entirely in Spanish, that informs the public, that accompanies people from their first exposure to a concrete contribution and that produces citable evidence and content along the way.

In short: if we build in Spanish the infrastructure that is missing (training, evidence, network and institutional presence) and make it easy to enter, we will produce the people and the knowledge that make AI safer for those who live in this language.

3. The logic model

Inputs

Activities

Organized according to the blocks of What we do:

• Train: the AI safety Seminar, with BlueDot Impact, and the AGI Strategy Course, cohort-based and with no technical requirements.
• Connect: a calendar of open sessions and social meetups on Luma, and ongoing discussion on the iaS Discord.
• Research: Paper digest, translation of materials and maintenance of the bilingual glossary.
• Communicate: a weekly newsletter, videos and social media posts that bring the evidence and news of the field into Spanish.
• Chapters: iaS Valencia as the first chapter, and a procedure for starting new chapters.
• Advocacy: taking part in public consultations and in institutional forums.

Outputs

Figures from the iaS CRM as of September 19, 2026; team members are not counted in the recurrence figures.

Expected outcomes

1. An evidence base in Spanish. iaS documents how AI systems fail in this language and takes that evidence to regulators, developers and media, who cite it as a source.
2. A pathway for Spanish-speaking talent. People who go from their first session to training, contributing and working in AI safety (policy, research, industry or civil society), without language being a barrier.
3. Greater literacy and public debate in Spanish. AI safety enters the conversation of universities, companies and media across the Spanish-speaking world, and those who take part in it act with better judgment.
4. A network of chapters. The model proven in the online sessions is replicated in other cities and countries of the Spanish-speaking world, starting with Valencia, with a shared brand, materials and procedures.
5. Institutional partnerships. Universities, administrations, civil society organizations and companies turn to iaS for training, collaboration and dialogue on AI safety.

Impact

The development and deployment of AI is safe and fair in Spanish too, and the people who live in this language take part in the decisions that shape it.

How we measure it

iaS plans to measure the change attributable to its activities through entry and exit forms at each session (for example, the hours each person spends on AI safety before and after), testimonials from people who have changed course thanks to them, and individual follow-up in the CRM.

4. Evidence and validation

The problem is documented. Multilingual studies such as M-ALERT and the ODESIA comparison table (see section 1) show differences in behavior and effectiveness across languages; iaS does not claim that models are less safe in Spanish in every case, but that these differences exist and vary by risk category.

There is demand. Since February 2026, the open events have brought together more than 250 distinct people, with 446 people who have confirmed their registration at least once, around one hundred confirmed registrations per session from May to July and an average rating of 4.56 out of 5, with 95% of the ratings at 4 or 5.

The format retains part of the audience. About 60 people have attended two or more events (around a fifth of those who have come) and 30 have attended three or more, and attendance over confirmed registrations at the online sessions from February to July has ranged between 35% and 54%.

What we do not yet know. iaS has been running since February 2026, a period too short to observe career transitions; we do not yet measure the counterfactual change (what each person would have done without iaS), recruitment through direct invitations from Luma has not produced attendance, and the first chapter, Valencia, is at an early stage and does not yet offer results to evaluate.

5. Scaling logic

The model produces a cycle at the center of the theory:

• More sessions and more chapters, more people entering the field in Spanish.
• More people, more data of our own on what the public wants and what retains it, and more citable content and evidence.
• More evidence and more data, more credibility with institutions and funders.
• More credibility, more resources for materials, programs and chapters.
• More resources, more chapters interested in joining, and the cycle repeats.

This cycle works because AI safety in Spanish combines two needs that no local group covers on its own: presence in each city, which makes in-person meetings and local roots possible, and an infrastructure shared across the whole Spanish-speaking world, which makes possible the common language, the glossary, the evidence, the brand and dialogue with institutions and funders; Spanish is, in this sense, the infrastructure that unites the five continents.

For a new chapter, iaS provides the brand, the materials, the Discord, the registration form and the management of funding applications, so that those who lead it do not have to pay for events out of their own pocket or start from scratch; in this first phase, the brand, the funding and the attendee data are managed centrally from iaS, and the model will be reviewed according to the results in Valencia.

The role of iaS is to catalyze and sustain this cycle: to maintain the quality of the central infrastructure, to support chapters through their difficulties and to keep the mission at the center as the organization grows.