A project reflection by SOMOS Civic Lab, supported by the Pulitzer Center through the South to South (S2S) AI Accountability CoLab microgrant.
The public servants closest to citizens are usually the last to be trained on the technology reshaping their work. Decisions about how AI models behave are made by a handful of companies far from Latin America. The consequences arrive at the municipal level, where a communications officer must decide in minutes whether a viral audio clip is real and what the government should say about it.
Our project was built for that officer. With the Pulitzer Center microgrant and in partnership with Aceleradora de Ciudades, a network supporting 20 Mexican municipal governments, SOMOS Civic Lab delivered a five-session intensive program on ethical AI and electoral integrity to an average of thirty civil servants per session, drawn from more than ten cities and the Instituto Nacional Electoral (INE). A second component, a Critical AI Literacy workshop co-designed with indigenous programmers in Brazil, is in advanced planning. Both follow the same architecture of critical AI literacy, hands-on evaluation, and collective deliberation.
Where the project sits in the AI accountability landscape
Latin America is entering one of the most intense electoral periods in its history, with Mexico's 2027 elections approaching. The systems that will shape these elections are evaluated, if at all, in the laboratories that develop them, in languages and contexts that are far removed from those of a Mexican municipality. Our project is relevant to accountability because it shifts the focus of AI evaluation. Rather than leaving this task to the companies, it is transferred to the institutions and communities that are responsible for the technology's failures, equipping them to perform this task effectively.
What the grant and the network made possible
The grant gave us the capacity to move from concept to delivery. The Pulitzer Center's grantee network gave us something we did not anticipate when we applied. For most projects, a funder's network sits adjacent to the work. In ours, they became instructors. Chequeado, the Argentine fact-checking organization and a fellow grantee, trained our participants to anticipate predictable false narratives, prioritize the cases that could cost citizens their vote, and fill information gaps in formats people actually see. WITNESS shared its rapid-response channel, which connects fact-checkers with deepfake detection experts. Cazadores de Fake News opened its methodology for investigating coordinated influence operations. Meedan presented the landscape of disinformation in Mexico during the 2024 elections and highlighted the value of coalition-building. INE closed the program by presenting its own verification methodology and inviting municipalities to work with it directly. A grant network became standing infrastructure.
How journalism was useful and where it fell short
Journalism gave the curriculum its evidence base. Research by Eon Institute on generative AI in Mexico's 2024 elections reviewed nearly 6,000 news articles and confirmed 44 real cases, from cloned candidate voices to deepfake investment scams to non-consensual intimate imagery targeting women candidates. Reporting also showed the constructive face of the same technology. A collaboration between Meedan, INE, AFP Factual, Animal Político, and Telemundo put more than 90 fact-checks in front of citizens through WhatsApp during the 2024 election. And the working methods of journalism such as verification before publication, refusing to amplify while warning, and speaking with a single voice directly influenced the anti-disinformation protocols our participants drafted for their own municipalities in the final session.
The limitations were just as instructive. Localized, Spanish-language reporting on AI accountability remains scarce. And reporting, on its own, does not change institutional behavior. In our original proposal we imagined curating Pulitzer Center-supported stories as reading material. What worked was convening the people behind the reporting and converting their methods into exercises, simulations, and working documents, like a simple electoral disinformation protocol for municipalities based on a Whatsapp viral disinformation exercise. For this, we simulated a disinformation WhatsApp campaign on screen for participants that was produced with AI (the screen itself and text in WhatsApp) and we asked participants to apply a simple disinformation protocol to take action on a hypothetical day of elections. This helped them understand that a protocol can be simple even if the tech is sophisticated.
Challenges along the way
The public conversation about AI was our first obstacle. It swings between inflated promise and inflated fear, and neither mood produces institutional readiness. We responded with evidence and rehearsal, giving participants a measured picture of the risk and concrete practice in responding to it, so the technology became something to manage rather than something to dread. In Brazil, the challenge was of a different kind. Co-designing a tailored workshop with Indigenous partners takes time, and we chose to shift our timeline rather than compress theirs.
What comes next
The Brazil workshop is the next test of the blueprint, with three sessions planned on critical AI literacy built with and for Indigenous programmers, outside the electoral frame, where participants define for themselves which harms matter and how the systems in their lives should be evaluated. In Mexico, the protocols our participants drafted now travel toward 2027, and we intend to reconvene the cohort as the elections approach.
In our first session, Amaralina Xavier of Meedan left the group with a charge: “We should not reject AI. We should want to be part of the decisions about how it will be used and how it will be regulated in our countries.” She was speaking to civil servants, but she was describing a region. The seat at the table where AI is evaluated and governed will not be offered to Latin America. Projects like this one exist to take it.
