From Algorithm Adoption to Institutional Oversight: A New Layer of AI Governance for Latin America
- Isabel Velarde
- Apr 7
- 12 min read
Updated: May 19
AI governance in Latin America is rapidly evolving, but the region still faces critical gaps in institutional capacity to ensure algorithmic accountability, transparency, and equitable decision-making.
Someone in your organization approved a system that decides who gets credit, who gets hired, and who gets access to public services in this region. Almost certainly, no one asked where that system was trained, on whose data, and for which market. Almost certainly, no one knows what it cannot see. That is not a technology problem. It is a governance failure, and it is happening at scale, today, inside institutions that would never accept equivalent opacity in a financial audit.
The algorithm was not designed for the Latin American market. It was designed for a different one, exported as a service, and adopted without examination.
Position determines everything. Raúl Prebisch was not a revolutionary. He was a central banker who had seen the numbers. In 1950, presenting before the United Nations Economic Commission for Latin America with the restraint of a technician, he identified a structural regularity that prevailing theory was not prepared to confront: the terms of trade between the industrialized center and the exporting periphery did not converge over time. They diverged. The center designed, standardized, and captured margins. The periphery supplied raw materials and received a fraction of the value generated. Position within a value chain, Prebisch understood, determines everything.
Seventy-five years later, the commodity has changed. The structure has not.
The new commodity: data. Every credit inquiry, every job application submitted through a digital portal, every search from a smartphone in São Paulo or Lima or Bogotá produces information that flows into processing infrastructures located primarily in the United States, China, and increasingly Europe, in jurisdictions that bear no legal or political responsibility toward the individuals whose behavior generated that data. It returns, transformed, as a decision. What appears as a sequence of technical operations is, in fact, the institutionalization of a separation: economic reality is produced in Latin America; the criteria used to interpret it are defined elsewhere.
The raw material leaves. The verdict returns.
Scholars such as Couldry and Mejias (2019, 2024), Mohamed, Png, and Isaac (2020), Birhane (2020), and Kwet (2019) have analyzed this asymmetry as forms of digital colonialism and data extractivism. Building on these analyses without reproducing them, what is observable in Latin America is a specific structural pattern: data operates as the raw material, models operate as the manufactured product, and the margin is captured where processing takes place, almost always outside the region. China is not only a processor in this arrangement: it refines critical minerals and simultaneously exports AI models to the region through platforms developed by Alibaba, Tencent, and ByteDance. The center has multiplied. The structure has not changed.
Structural invisibility
These models were built on data from economies where formal employment is the rule and financial histories are dense. When deployed in a region where, according to the International Labour Organization's 2025 Labor Panorama (the most recent data available at time of writing), 46.7 percent of employment is informal, rising to 56 percent among young people, they do not merely struggle to assess the informal economy. They structurally exclude it before any analysis takes place. This is not a performance issue. It is a design limitation.
The result is not that informal workers are classified as high risk. They are not classified at all. They fall outside the system's field of vision. There is no record of that non-evaluation, no possibility of appeal against a decision that was never formally issued. There is only absence, and that absence, because it generates no signal, is interpreted as confirmation of the original design. Invisibility becomes self-reinforcing.
This is the structural condition that defines significant segments of Latin America's digital economy: productive activity that exists in reality but has no representation in the digital maps automated systems use to make decisions. The merchant who operates in cash with a flawless local reputation is invisible to a model that requires formal deposits. The professional whose career shows interruptions for care work is penalized by systems that read those pauses as instability. The asset owner whose property lacks formal registration appears nonexistent in valuation models. Economic value exists. What does not exist is its translation into the format that the system recognizes as valid evidence.
This structural invisibility is precisely what the concept of Pre-Evaluative Exclusion identifies (Velarde, 2026, DOI: 10.5281/zenodo.19665795): a class of algorithmic harm that operates before evaluation occurs, by determining who or what becomes admissible to evaluation in the first place.
Gender as a structural mechanism
Gender is the lens through which this exclusion is most precisely reproduced. The World Bank's Global Findex Database documents that women in Latin America remain significantly underrepresented in formal financial systems, a gap not explained by income or education but by structural factors that automated systems, calibrated on data from other markets, systematically misread as risk indicators. CEPAL's Social Panorama 2024 documents a labor force participation gap between men and women exceeding twenty percentage points, driven by the concentration of unpaid care work on women. A talent-screening system trained on historical hiring data from an industry where women have been systematically underrepresented does not discriminate in any legally legible sense. It learns that advancement profiles are profiles women have historically been prevented from building, and replicates that learning at scale.
The 2018 Amazon case is the most documented instance of this mechanism. According to Reuters reporting, a model trained on ten years of the company's own hiring data had learned to penalize signals associated with female candidates, not through any explicit variable, but through the historical data it had been trained to replicate. A study published in Science in 2019 by Obermeyer, Powers, Vogeli, and Mullainathan documented the same mechanism in health care: an algorithm managing care for millions of US patients systematically underestimated the needs of Black patients because it used historical cost as a proxy for health need, without accounting for the barriers that had historically constrained their access to care. The exclusion was not in the variable. It was in the data the variable reflected.
The bias is not in the code. It is in the history the code was trained to repeat.
This insight, developed extensively by Buolamwini and Gebru (2018), O'Neil (2016), Eubanks (2018), and Noble (2018), reframes algorithmic discrimination from a technical problem into a question of which historical patterns institutions choose to scale.
A question the room was not prepared to process
In July 2019, at the Inter-American Development Bank's Made in the Americas Global Digital Services Summit in Buenos Aires, a Chilean executive interrupted a conversation about credit scoring platforms with a question the room was not prepared to process. Have you noticed, she said, that our evaluation systems are biased? She was not referring to explicit discrimination but to something structurally deeper: algorithms designed in Silicon Valley, calibrated for markets of dense formal employment, applied uniformly to economies where nearly half the labor force operates outside the formal system. When providers were confronted with the exclusions their systems produced, the response was consistent: the model reflects the reality of the market. It was not the Latin American market. What was presented as neutrality was the projection of an external standard onto a structurally different context.
That conversation was not exceptional. The algorithm arrives packaged from external providers, carrying embedded assumptions about creditworthiness, employability, and eligibility, defined, tested, and optimized elsewhere. The institutions that implement them, almost without exception, never ask what those assumptions are. They ask how much the system costs, how quickly it can be deployed, and what return it will generate. The criteria, the very definition of who is included and who is excluded, arrive by default.
Accelerated consumption, minimal governance
The Latin American Artificial Intelligence Index, ILIA 2025, published by CEPAL and Chile's National Center for Artificial Intelligence, documents the imbalance with precision. Latin America represents 6.6 percent of global GDP but receives only 1.12 percent of global AI investment, six times less than the global average relative to GDP per capita. Yet the region accounts for 14 percent of global visits to AI solutions, despite representing only 11 percent of global internet users. Accelerated consumption. Minimal production. No proportional governance.
Prebisch called it the deterioration of the terms of trade. Today, Latin America faces a structurally similar question: the region supplies the raw material of the digital economy and receives, packaged as innovation and priced as a service, the criteria that govern its own participation in the market. What is being exchanged is no longer goods. It is judgment itself.
The physical infrastructure of this ecosystem makes the double dependency concrete. According to UNCTAD's Digital Economy Report 2024, data centers consumed 460 terawatt-hours of electricity in 2022, equivalent to France's total consumption, projected by the International Energy Agency to reach 1,000 terawatt-hours by 2026. Latin America supplies not only data but the minerals that make the hardware possible. According to the U.S. Geological Survey Mineral Commodity Summaries (2024), Chile is the world's leading lithium producer, while the Democratic Republic of the Congo accounts for the majority of global cobalt production. China dominates the refining stage of both materials and most rare earth elements essential to semiconductor manufacturing. The World Bank's Minerals for Climate Action report projects substantial increases in demand for these critical materials in the coming decades. The region extracts the material foundation of the digital economy, then pays recurring license fees to access the intelligence built on that substrate. Mineral contribution. Digital contribution. Surplus captured elsewhere.
Delegation without oversight
The governance gap is institutional, not informational. The default response, calling for more and better data to capture informal economic activity, misdiagnoses the problem. The deficit is not in the data. It is in the accountability structures that transform data into decisions. Boards of directors across the region approve budgets, vendors, and implementation timelines. They rarely approve criteria. The definition of who is included and who is excluded arrives embedded in imported systems and, once deployed, operates continuously and at scale without any deliberative body examining the values encoded or the populations made invisible.
The absence of an explicit decision does not imply neutrality. It implies delegation without oversight, and delegation without oversight is a fiduciary failure.
Three levels of action
The response is not autarky. Latin America neither needs nor can build the entire AI stack internally. The response is to build a new layer of institutional capacity, the capacity to audit, condition, and when necessary refuse the criteria embedded in systems before they are deployed. This is the layer that transforms AI adoption into accountable AI governance.
At the board and institutional level: before deploying any automated system with material impact on access to credit, employment, or public services, institutions must commission independent algorithmic audits establishing which populations the system cannot evaluate, what proxy variables it uses and what they actually measure, and who bears accountability when errors produce harm. This is not a technical exercise. It is a governance obligation equivalent in character to financial due diligence, and it should be treated as such by audit committees. The Institutional AI Deployment Review (IADR), introduced in Velarde (2026, DOI: 10.5281/zenodo.20217979), offers a structured framework for boards, executive leadership, and procurement functions to conduct this kind of pre-deployment review.
At the regulatory level: governments should require, as a condition of market access, that providers of high-impact automated systems disclose training data provenance, demonstrate performance across informal-economy segments, and establish local accountability mechanisms. Brazil's Lei Geral de Proteção de Dados already establishes the right to review decisions taken solely on the basis of automated processing. Colombia's framework on financial and credit information requires lenders to demonstrate they considered factors beyond automated scoring outputs. The legal foundations exist. Enforcement infrastructure must follow.
At the data infrastructure level: the region must invest in building representative datasets that reflect its own economic reality, including informal labor, cash-based commerce, and non-registered assets, so that the next generation of models can be trained on evidence that corresponds to the populations they will govern. Brazil's Nubank has demonstrated that local adaptation is possible: by incorporating alternative data such as utility payments and rental history, it has extended credit access to segments that traditional scoring systems rendered invisible, though not without raising privacy questions that deserve equivalent governance attention. Scaling this approach requires public investment, inter-institutional coordination, and the recognition that data infrastructure is as strategic as physical infrastructure.
Regulation is beginning to move
The European Union's Artificial Intelligence Act, published in July 2024 and in force since August 1, 2024, classifies credit evaluation, employment decisions, and access to essential services as high-risk applications subject to mandatory conformity assessments, human oversight, and the right of affected individuals to be informed that an automated system evaluated them. Full enforcement obligations for most high-risk systems apply from August 2026. The Act does not apply directly to Latin American jurisdictions, but it does apply to the external providers whose systems the region deploys, and its governance conditions will migrate as a condition of continued access to the European market. The direction of regulatory convergence is unambiguous. The pace remains slower than the adoption of the systems it seeks to govern.
Building the new layer
Prebisch's question was about who captures the surplus generated between center and periphery. The question today is about judgment: who defines it, who applies it, and who answers for its consequences. The algorithm deciding who gets credit in Lima or who gets hired in Bogotá is not a neutral instrument. It is a political choice, embedded in systems developed outside the region, applied as a service, and governed, in most cases, by institutional structures that have not yet been built to oversee it.
Every board member who approves a vendor contract without reviewing the system's evaluation criteria is making a governance choice, whether deliberately or by default. Every regulator who has not yet required disclosure of training data provenance is granting an exemption, by default. Every institution that deploys an automated decision system without auditing its exclusions is defining who is visible in this market, whether they intend to or not.
The opportunity is to make these choices deliberate rather than implicit, accountable rather than delegated. Latin America does not need to produce its own AI stack to govern AI responsibly. It needs to build the institutional layer that transforms imported algorithms into governable systems, with criteria reviewed by its own boards, regulators, and oversight bodies.
Prebisch looked at the numbers and refused to mistake a structural arrangement for a natural law. That is the same intellectual discipline that Latin American institutions, public and private alike, are now invited to apply. The configuration he described took decades to name. This one is being named now. The time to build the new layer is before it becomes entrenched, not after.
References
1. CEPAL and Centro Nacional de Inteligencia Artificial de Chile (CENIA). Latin American Artificial Intelligence Index, ILIA 2025. Santiago: CEPAL, October 2025.
2. International Labour Organization. Panorama Laboral 2025: América Latina y el Caribe. Geneva: ILO, December 2025.
3. United Nations Conference on Trade and Development. Digital Economy Report 2024: Shaping an Environmentally Sustainable and Inclusive Digital Future. Geneva: UNCTAD, July 2024.
4. World Bank. Global Findex Database 2021: Financial Inclusion, Digital Payments, and Resilience in the Age of COVID-19. Washington, DC: World Bank, 2022.
5. CEPAL. Panorama Social de América Latina y el Caribe 2024. Santiago: CEPAL, 2024.
6. Obermeyer, Z., Powers, B., Vogeli, C., and Mullainathan, S. Dissecting racial bias in an algorithm used to manage the health of populations. Science, 366(6464), 447–453. October 2019.
7. Dastin, J. Amazon scraps secret AI recruiting tool that showed bias against women. Reuters, October 10, 2018.
8. European Parliament and Council of the European Union. Regulation (EU) 2024/1689 (Artificial Intelligence Act). Official Journal of the European Union, July 12, 2024. In force August 1, 2024; high-risk system obligations applicable from August 2026.
9. Brasil. Lei nº 13.709/2018 (Lei Geral de Proteção de Dados Pessoais, LGPD). Brasília, 2018.
10. Colombia. Ley 2157 de 2021, que modifica y adiciona la Ley 1266 de 2008. Bogotá, 2021.
11. Prebisch, R. The Economic Development of Latin America and its Principal Problems. New York: United Nations Department of Economic Affairs, 1950.
12. Couldry, N. and Mejias, U. A. The Costs of Connection: How Data Is Colonizing Human Life and Appropriating It for Capitalism. Stanford: Stanford University Press, 2019.
13. Couldry, N. and Mejias, U. A. Data Grab: The New Colonialism of Big Tech and How to Fight Back. Random House, 2024.
14. Mohamed, S., Png, M.-T., and Isaac, W. Decolonial AI: Decolonial Theory as Sociotechnical Foresight in Artificial Intelligence. Philosophy & Technology, 33(4), 659–684. 2020.
15. Birhane, A. Algorithmic Colonization of Africa. SCRIPTed, 17(2). 2020.
16. Kwet, M. Digital Colonialism: US Empire and the New Imperialism in the Global South. Race & Class, 60(4), 3–26. 2019.
17. Buolamwini, J. and Gebru, T. Gender Shades: Intersectional Accuracy Disparities in Commercial Gender Classification. Proceedings of Machine Learning Research, 81: 1–15. 2018.
18. O'Neil, C. Weapons of Math Destruction: How Big Data Increases Inequality and Threatens Democracy. New York: Crown Publishing Group, 2016.
19. Eubanks, V. Automating Inequality: How High-Tech Tools Profile, Police, and Punish the Poor. New York: St. Martin's Press, 2018.
20. Noble, S. U. Algorithms of Oppression: How Search Engines Reinforce Racism. New York: NYU Press, 2018.
21. U.S. Geological Survey. Mineral Commodity Summaries 2024. Reston, VA: U.S. Geological Survey, January 2024.
22. World Bank. Minerals for Climate Action: The Mineral Intensity of the Clean Energy Transition. Washington, DC: World Bank, 2020.
23. Velarde, I. Who the Algorithm Doesn't See: A Practitioner Audit for Pre-Evaluative Exclusion in High Informality Economies (Latin America). Zenodo, 2026. DOI: 10.5281/zenodo.19665795
24. Velarde, I. The Governance Gap Behind AI Deployment in High-Informality Economies: Board Oversight, Procurement Asymmetry, and Operational Risk. Zenodo, 2026. DOI: 10.5281/zenodo.20217979
About the author
Isabel Velarde is Executive Fellow in AI Governance, Responsible AI, and Digital Transformation at the Digital Growth Collective, London; Senator and Country Chair for Peru, and President of the Gender Equality in Technology Committee, at the World Business Angels Investment Forum, affiliated with the G20 Global Partnership for Financial Inclusion; and Founding Partner at LACBAN. She is the author of practitioner working papers on AI governance and algorithmic accountability in Latin America, published on Zenodo (DOIs: 10.5281/zenodo.19665795 and 10.5281/zenodo.20217979).
Comments