Blind by design: what AI exposure indices miss about Global South labor markets and why it matters for the future of work
Blog l Ramiro Albrieu, June 2026
Last week I went to Cape Town. I was there for a project I'm collaborating on that tries to see with fresh African eyes how AI is transforming — and might transform — jobs. On the flight over, I spent a few hours reviewing the material I had to present: a curated review of the leading indices that measure AI exposure across occupations and what they tell us about the North–South divide.
The takeaway from those indices is consistent: high-skill workers are the ones exposed and the remainder of the labor market (let's call them lower-skill workers) barely enters into the picture [1]. High-skilled workers account for 7% to 12% of total jobs in low- and lower-middle-income countries, and almost 45% in the Global North [2]. The conclusion is clear. The story seems settled — and it shapes much of the global debate on the future of work.
Then I landed in Cape Town.
Walking through the city I kept seeing the same thing: a street vendor checking prices with a voice assistant. A driver routing with an AI-powered app. A small trader photographing stock on a smartphone, letting an app track what sold. A domestic worker managing client schedules through a messaging bot. Standing there, I realized that I've been watching the same scene for years but closer to home, in Latin American labor markets. The plumber who uses ChatGPT to write up a quote and coordinates with clients over WhatsApp voice messages. The seamstress running her business's Instagram from her phone, using AI-generated captions. The street food vendor who checks ingredient prices with a bot and keeps her accounts on a free AI-powered app. By most standard exposure indices, none of these Latin American and African workers are exposed to AI. By any observational account, they absolutely are.
Something in the standard picture does not add up.
This apparent paradox reminded me of Alejo Carpentier's El reino de este mundo [3]. Carpentier observed that what strikes the outside observer as marvelous, paradoxical, anomalous — the extraordinary bleeding into the everyday — is simply the ordinary texture of experience in this part of the world. A lower-skill worker using frontier AI in their daily job looks, from the vantage point of the standard exposure literature, like a piece of magic realism. But from a Global South perspective, it is a reality. A reality that, with the help of proper actions to shape change, can make AI transformative in a truly global sense.
Why do traditional exposure indices treat these realities as anomalies and fail to detect the immense possibility that comes with them? Let's look at the tasks workers perform in the busy streets of Cape Town or Lima. These are labor markets characterized by low occupational specialization, which means that every worker is also her own manager, accountant, web designer, and admin. In the Global North, these peripheral tasks can be contracted away to highly specialized segments of the labor market. In the Global South, in contrast, many peripheral tasks don't have dedicated occupations (or their markets are very thin [4]); workers must absorb them as part of their daily job. Thus, in Africa and Latin America the "jack-of-all-trades" job — or todero [5] — is not an anomaly; it is the modal worker. Labor market organization differs between the North and the South; in the latter, workers spend significant time on peripheral tasks, preventing specialization and decreasing productivity.
And that's where AI enters the picture: AI is outperforming humans at precisely these peripheral tasks. The most common AI-assisted work activities involve information work: the creation, processing, and communication of information [6]. That maps directly onto the ample space of administrative, commercial, financial, and logistical tasks. If AI is most powerful exactly where these workers' burden is heaviest, the exposure indices -built on the assumption that occupations contain the same tasks everywhere- are missing a very important part of the story.
Actual AI usage tells us the opportunity is there. To test this — whether AI is skill-biased or skill-equalizing — in Sur Futuro we partnered up with CEDLAS, Di Tella University, and others to run a randomized experiment with more than a thousand participants. The finding was clear: lower-skilled workers get the largest productivity gains from generative AI, closing roughly 75% of the performance gap with higher-skilled workers [7]. That's the multiplier hiding in plain sight.
How to push this further and make it happen by design, in a more generalized and productivity-oriented way? The starting point is getting better information. We don't know with enough precision what AI tools low- and medium-skill workers in the Global South are actually using — the voice assistants, the chat-based tools, the smartphone apps that are the real interface between these workers and the technology, and that don't appear in any existing exposure index. And we know even less about what tasks these workers actually perform in their jobs. The granular, occupation-level task data that informs AI exposure measurement was built for rich-country labor markets. For Africa and for Latin America, that data — at the level of what a plumber, a market vendor, or a domestic worker actually does hour by hour — largely doesn't exist.
If we can gather real data on both the most-used AI applications and the task content of jobs, a rich and previously unknown design space will open up. On the technology side, start imagining products that are not there yet: mobile-first AI tools built for real conditions of use in the South — low connectivity, local languages and registers, interfaces that don't assume cultural capital that isn't always there. On the skills side: training that goes beyond surface-level courses and develops workers' real capacity to evaluate, question, and adapt AI outputs to concrete situations in their trades — not generic reskilling recommendations, but specific answers to specific questions: what does a plumber, a seamstress, or a food vendor need to learn to work better with AI? That's the design space for a future of work that actually includes the Global South.
The bias embedded in the standard exposure indices is not only geographic (that usage data comes almost exclusively from the US) or occupational (that only office-oriented, digitally-instrumented jobs leave a trace in the datasets). It is ontological: the very categories used to define what a "job" is, what "tasks" it contains, and therefore what "AI exposure" means, were built for a particular kind of economy. They don't fail to see the street vendor or the todero because of a data gap. They fail to see them because their conceptual architecture was never designed to. This is what Carpentier was pointing at: the anomaly doesn't lie in the reality. It lies in the lens.
I returned to Buenos Aires with one idea in mind: that AI can transform lower-skill jobs, and with that bring more productivity and wellbeing to Global South workers. Building a truly transformative AI for the Global South — and a future of work that works for everyone — means building new lenses, and then acting on what they reveal.
[1] The two most cited methodologies are the IMF's AI Occupational Exposure index (Felten, Raj & Seamans 2021; Pizzinelli et al. 2023) and the ILO's Occupational Exposure Index (Gmyrek, Berg & Bescond 2023, revised 2025). Both show a consistent North–South gap: 34% of employment in high-income countries falls in occupations with some AI exposure; in low-income countries, 11%. [2] ILO Modelled Estimates, Nov. 2025. High-skill occupations (ISCO-08 groups 1–3) represent 6.9% of employment in low-income countries, 11.6% in lower-middle-income countries, and 44.9% in high-income countries. [3] Alejo Carpentier, El reino de este mundo (1949). The argument appears in the novel's prologue, where Carpentier develops the concept of lo real maravilloso — the idea that what the outside observer experiences as extraordinary is, in this part of the world, simply the texture of the everyday. [4] Indeed, according to Bandiera et al. (2022), high-income countries have more than four times more occupation types than the poorest countries. [5] The term todero — broadly meaning "one who does everything" — is used across several Latin American countries to describe workers who operate in low-specialization labor market environments, taking on a wide range of tasks beyond their core trade. Eduardo Galeano evokes the phenomenon in Las venas abiertas de América Latina. A recent systematic analysis with Peruvian data can be found in Atencio-De León et al. (2025). [6] Tomlinson et al. (2025), analysis of 200k Bing Copilot conversations classified against ONET work activities; Chatterji et al. (2025), analysis of 1.1M ChatGPT interactions.* [7] Cruces, Fernández Meijide, Galiani, Gálvez & Lombardi (2026). NBER Working Paper 34851. https://www.nber.org/papers/w34851 [8] Massenkoff, M., Lyubich, E., Sacher, S., Hitzig, Z., Zhang, S., Heller, R. & McCrory, P. (2026). Anthropic Economic Index report: Cadences. Anthropic, June 26, 2026. https://www.anthropic.com/research/economic-index-june-2026-report. The report documents two compounding biases in available AI usage data: a geographic bias (microdata exists only for the US; country-level aggregates are available but don't identify who is using AI, for what task, or in what occupation) and an occupational bias (physical occupation categories — transportation, construction, food service — are under-represented in Claude sessions and in the survey).
