Opinion – Cognitive Sweatshops: AI Labor and the US-Latin America Divide
Franco Laguna Correa
Classical dependency theory described a world economy organized around a core that manufactures high-value goods and a periphery that supplies raw material and cheap labor. Regardless of formal political sovereignty, the terms of trade structurally favored the core. Contemporary scholarship on the political economy of AI has extended this framework to data and cognition. Nick Couldry and Ulises Mejías propose the concept of “data colonialism,” which frames the extraction of data and labor from the Global South as a continuation of historical colonial appropriation. In this account, digital platforms have replaced plantations or mines as the primary medium of extraction. A comparative study of data workers in Argentina, Brazil, and Venezuela confirms that most flows of AI data work reproduce this legacy pattern, connecting Latin America to corporate clients based in the United States and Europe.
Get E-International Relations delivered to your inbox, free of charge. As you sign up, consider becoming a paid subscriber, or make a donation, to support our work.
Parallel literature on algorithmic sovereignty argues that control over data infrastructure, compute, and the standards governing both is concentrated in a handful of states and firms. As a result, Global South governments have limited capacity to set the terms of their own participation in AI production. These frameworks give the “cognitive sweatshop” thesis its international relations register. The phenomenon is not simply exploitative employment; rather, it is a structural feature of how value, cognition, and dependency are currently distributed across the hemisphere.
Some governments in the Global South, including Brazil, have begun responding to this asymmetry by investing in state‑funded foundational models trained on domestic languages and data pools. This approach is sometimes described as building sovereign AI stacks. Whether such initiatives can alter the underlying labor architecture, or will simply relocate the same wage‑arbitrage logic under domestic ownership, remains an open question. The point here is that AI infrastructure has already become a site of soft power and geopolitical maneuvering, not merely a technical or commercial matter.
The mechanism binding cognition to political economy is wage arbitrage operating across a value chain whose commanding heights sit with a small set of firms. Recruitment intermediaries advertise Latin American annotation labor to United States clients at roughly a third of comparable domestic wages, explicitly marketing the discount as the product’s selling point. This is not incidental cost‑saving; it is the organizing logic of the industry. Global survey of digital labor platforms find that identical tasks, performed for the same international clients, are compensated at markedly different rates depending on the worker’s country of residence. This differential reflects worker competition and platform targeting criteria such as language and location, rather than any difference in task complexity.
The value generated by this labor—improved model safety, more capable products, and ultimately corporate revenue and market capitalization—accrues almost entirely to firms headquartered in the United States. Meanwhile, the compensation returned to the periphery is calibrated to local cost of living rather than to the labor’s contribution to a global product. This asymmetry is the economic core of what dependency theorists would recognize as unequal exchange: labor of comparable cognitive intensity is priced differently depending on its geographic origin, and the resulting surplus is repatriated northward.
A second, less visible dimension of this architecture concerns currency. Much of this labor is compensated in United States dollars, which workers in inflation‑prone or currency‑restricted economies actively seek out as a hedge against domestic monetary instability. This gives the exchange a dual character. It is simultaneously a labor transaction, in which cognition and attention are sold below their marginal contribution to the client firm’s output, and a currency transaction, in which access to hard currency substitutes for the absence of stable domestic financial instruments. Both dimensions deepen dependency rather than resolving it, since neither builds durable institutions, transferable skills, or bargaining power within the worker’s home economy.
It is useful to locate this exchange precisely within the value chain. Large United States technology companies and the frontier laboratories they fund rarely contract with individual annotators directly. Instead, a layer of intermediary firms and business‑process outsourcers sits between capital and labor. These intermediaries absorb regulatory risk and enforce productivity standards while formally bearing the title of employer. This intermediation lets lead firms benefit from cheap, flexible cognitive labor while remaining legally and reputationally insulated from the conditions it entails. In international relations terms, the intermediary layer functions much like the local elites of classical dependency accounts: it translates the core’s demands into terms the periphery can supply, captures a share of the surplus for itself, and leaves the worker at the base of the hierarchy with the least bargaining power of all.
Brazil illustrates this architecture with unusual clarity because its macroeconomic profile differs from crisis‑driven cases like Argentina and Venezuela, which are often used to explain why Latin Americans turn to platform data work. Brazil is the largest economy in Latin America, one of the world’s largest exporters of outsourced computing services, and a major producer of information technology graduates. Yet it remains among the most unequal countries in the world, with roughly two‑fifths of its workforce in informal employment.
One of the largest surveys of Brazilian data workers to date—477 respondents—found average monthly earnings from platform data work of approximately $112 for around 15.5 hours of work per week. This sum falls well short of estimated living costs in São Paulo, Rio de Janeiro, and Belo Horizonte, where over half of surveyed workers reside. Unlike in Argentina, where data work functions as a supplementary activity for relatively better‑off, internet‑savvy young men seeking dollar income amid currency controls, Brazilian data work is disproportionately taken up by structurally disadvantaged workers. Nearly two-thirds of respondents in one of the largest studies were women, many of them mothers relying on data tasks as their primary income precisely because the flexibility of remote micro‑tasking accommodates unpaid care responsibilities that formal employment does not.
This demographic profile matters because it shows that Brazil’s insertion into the global AI value chain does not follow a simple crisis‑migration logic, as in Venezuela, where data work has become something close to a national survival strategy amid economic collapse. Brazil’s participation is instead structural. It is a stable, populous, unequal economy with a large, educated, underemployed workforce supplying cognitive labor to United States technology firms as an ordinary feature of its integration into the digital economy, not as an emergency response to macroeconomic shock. This is precisely the condition dependency theorists describe as a durable, rather than transitional, peripheral position. A favorable exchange rate or a change in domestic policy is unlikely to dislodge this position on its own, because the underlying demand is structured by the pricing decisions of firms located elsewhere.
The cognitive sweatshop framework contributes a labor‑process dimension that purely macroeconomic accounts of dependency tend to omit. It is not sufficient to note that Brazilian workers are paid less for comparable output. The output itself is structured to extract sustained attention and perceptual judgment through fragmented, quota‑driven micro‑tasks that build no transferable expertise. The economic asymmetry described above therefore has a cognitive corollary. The same architecture that captures the surplus value of Brazilian labor also captures the worker’s attentional and interpretative capacity, without returning any of the professional development that would ordinarily justify sustained cognitive effort in a labor market.
Brazilian interviewees in the referenced study describe grievances specific to this arrangement. They work against foreign time zones that push shifts into early morning hours, and they experience isolation because minimal contact with same‑language peers compounds excessive cognitive fatigue. These are best understood not as incidental workplace friction but as the sensorial and social costs of servicing demand structured entirely around a client economy on the other side of the equator. In international relations terms, this is a dependency operating simultaneously at the level of macroeconomic value capture and at the level of the individual nervous system predominantly through digital eye strain.
The cognitive sweatshop describes a hemispheric arrangement in which the attention and judgment of young workers—concentrated in, though not exclusive to, Brazil—function as an input comparable to the raw commodities of earlier eras of dependency. This input is priced not by its contribution to the finished product but by the cost of living in the country where workers happen to reside. United States technology firms sit at the commanding heights of this arrangement, capturing the surplus generated by cross‑border wage differentials while leaving the underlying dependency undisturbed.
Addressing this will require more than firm‑level ethics codes. It calls for the kind of due diligence and supply‑chain regulation increasingly discussed in European policy circles, extended explicitly to AI data‑work chains. It also calls for domestic policies in Brazil and comparable economies that recognize and credential this labor rather than treating it as a residual, informal activity. Absent of such interventions, the cognitive sweatshop is likely to remain what it currently is: a stable feature of the hemispheric political economy, not a transitional labor‑market anomaly awaiting correction by growth alone.
Franco Laguna Correa holds a Ph.D. in Cultural Studies from UNC-Chapel Hill. He is a remote Research Associate at the University of Pittsburgh and works as an AI Evaluation Specialist at Blueprint Technologies. His scholarship focuses on the intersection of biopolitical dynamics, deterritorialization processes in global migration and digital humanism.