Workers across the region are training, annotating, and validating artificial intelligence models. Human workers contribute significantly to the annotation, evaluation, and alignment of many models. Yet this activity receives only 1.12% of global AI investment, while the AI talent gap relative to the global average has widened since 2022, notably marked by an accelerated loss of specialists.
General AI trainer roles have increased by 283% in global cross-border contracts in 2025 according to Deel.
The Essentials
- Deel observes a global increase of 283% in AI trainer roles across its cross-border contracts in 2025 (Deel 2026).
- The region covered by ILIA captures 1.12% of global AI investment (Latin American AI Index 2025).
- Ownership and control of data and intellectual property vary by project.
- The AI talent gap relative to the global average has widened since 2022, notably accelerated loss of specialists; the causal link to employment growth is not established by ILIA.
- The open question for the decade is whether Latin America can move from the status of labor supplier to that of actor in the AI value chain.
The Invisible Work That Powers Global AI
Behind every chatbot capable of understanding a nuance or avoiding gross bias lie thousands of hours of human labor. Annotating images, classifying responses, flagging problematic content, evaluating the relevance of a model output: these tasks constitute AI training work, and regional workers participate in this labor.
The Deel 2026 report on hiring in Latin America documents a rapid transformation of the regional labor market. AI-related roles—training, annotation, model evaluation—figure among the fastest-growing categories in international contracts. Mexico, Colombia, Argentina, and Brazil concentrate the bulk of this demand. The platforms recruiting typically originate from North America and Europe.
Latin America presents a rare combination: mastery of Spanish and Portuguese (two of the five most widely spoken languages in the world), proximity to the U.S. time zone, and digital infrastructure enabling remote work for a portion of the population and territories, although connectivity and accessibility deficits remain significant. The Deel 2026 report documents growth in AI Training positions in Chile of 209%, lower than the global 283%; this figure does not apply to all of Latin America.
The problem is that this workforce often works on data it does not own and under contracts whose technology transfer conditions vary.
283% Growth, 1.12% of Capital
The figure of 283% deserves to be read precisely. It measures the growth in job postings, not captured value. An AI Training job posting in Bogotá generates income for the worker and training data for the American company that orders it. The income remains partially in Colombia, while the data and models are controlled by the commissioning company according to contractual terms.
The ILIA 2025 indicates that the region represents 6.6% of global GDP and 1.12% of global AI investment. This ratio is not a statistical anomaly. It reflects a capital allocation structure where AI investment and company creation are concentrated in a few countries in the region; the causal effect on a subcontracting role is not established by the source.
More troubling still: the AI talent gap between Latin America and the global average has worsened since 2022. Regional workers perform notably data preparation, annotation, and verification tasks for AI, according to several functions identified in the EnCOre report.
This divide between employment and expertise is the hallmark of dependent integration: one enters the value chain, but from the bottom.
The Precedent of Raw Materials
Latin America has a long memory of this type of insertion. In the nineteenth century, it exported rubber, copper, saltpeter, coffee. Export revenues existed, but the rent, added value, industrial transformation, price control, remained in the metropoles. Economists called this the resource curse syndrome, or more precisely the thesis of peripheral dependence, developed notably by CEPAL economists in the 1960s.
The parallel with AI Training is not metaphorical. Training data is a resource. It has value because it is rare, contextual, linguistically and culturally situated. An annotation in Mexican Spanish is not interchangeable with one in Castilian Spanish. Content evaluation by a Brazilian worker who knows local social codes produces a different signal than that of a Philippine worker.
This specificity has value, enough for American companies to explicitly request it.
The commissioners often control the instructions and final use of outsourced annotation work, creating asymmetries in power and value. The annotator produces an input. The model trained with that input generates significant value. The annotator receives an hourly rate.
Economist Dani Rodrik defines “good jobs” as employment that creates transferable skills, decent income, and anchorage in a sector. AI Training work produces a new category of jobs whose quality remains to be established: it can build real skills or reproduce the structural precarity of platform work, in a technologically different form.
Actors Resisting This Pattern
The picture is not uniformly dark, and exceptions merit being named precisely because they show what is possible.
Argentina has developed an AI startup scene that produces not only annotators but founders. Companies like Tryolabs (computer vision), Mercado Libre (which built its own recommendation and fraud prevention models), or Satellogic (satellite imagery and geospatial analysis) show that it is possible to move from a peripheral market to building proprietary technological assets. These companies have raised capital, filed patents, and recruited local engineers in design roles, not merely execution roles.
Brazil has made a notable institutional choice: the Estratégia Brasileira de Inteligência Artificial, instituted in April 2021 and modified in July 2021, provides public funding for AI research, university-industry partnerships, and attention to AI governance, regulation, and ethical data use. For 2024-2028, the government launched PBIA, which provides for the development of models in Portuguese based on national data in order to strengthen AI sovereignty.
Colombia and Chile have launched accelerated machine learning training programs through their economic development agencies. The idea is to create pathways between workers entering the ecosystem through annotation and better-paid technical roles. These programs remain modest in scale, but their logic is sound: if AI Training can be an entry point rather than a glass ceiling, the dependence pattern can be partially broken.
These initiatives recall dynamics observed in other contexts of rapid technological deployment, such as the industrial integration of robotics in certain emerging countries, where local training allowed, in some cases, for moving up the value chain rather than remaining confined to assembly.
Growing in AI Employment Without Remaining a Subcontractor
For the decade 2025-2035, the central question is whether growth in AI Training employment in Latin America will produce a trajectory of skill-building and value capture, or whether it will remain limited to dependent integration.
Two trajectories are plausible, and current signals do not yet allow us to distinguish between them.
In the first, sustained demand for annotation and training work creates a critical mass of workers familiar with AI tools and logic. A portion of them, the most qualified, those with access to complementary training, move into technical roles: prompt engineering, model evaluation, fine-tuning, then engineering. Local startups capture this pool. Regional investors, or foreign funds attracted by a maturing ecosystem, begin to fund proprietary assets. Global investment progresses gradually.
This is the scenario of India in the 1990s-2000s: starting from IT outsourcing, it gradually produced engineers, then entrepreneurs, then global technology companies.
In the second trajectory, demand for annotation work remains structurally separated from model development. American companies optimize their protocols to reduce dependence on human workers; automation of annotation tasks could progress, potentially reducing the need for human labor. The pool of workers trained in annotation finds itself in a dead end: too qualified for unskilled jobs, not qualified enough for engineering roles that companies seek. The talent gap could continue to widen in this scenario. And governments, lacking coherent public policies, have not created the necessary pathways.
What would allow us to distinguish these trajectories in coming years is less the volume of jobs created than their nature. A positive signal would be an increase in academic AI publications from Latin American universities, or growth in regional patent filings in the sector. A negative signal would be wage stagnation in annotation roles despite volume growth, or accelerated automation of annotation tasks by the platforms themselves.
Policy choices matter. EBIA provides for measures likely to strengthen national capacities, notably research funding, partnerships, and improved data use. The challenge is that these policies require time and continuity, two resources that Latin American political cycles have historically struggled to produce.
Workers from Latin America participate in certain global chains for data production and AI system training. The true measure of its technological trajectory will be whether, by 2035, it figures more prominently in the executive committees of AI companies than in their subcontracting registries.
Sources
- Deel, Reporte sobre la contratación en Latinoamérica 2026
- Latin American AI Index 2025 (Universidad del Desarrollo, Chile)
- HiresLink, Data Q1-Q2 2026 on AI recruitment in Latin America
- Estratégia Brasileira de Inteligência Artificial, Ministério da Ciência, Tecnologia e Inovações, updated 2024
