Pennies Per Microtask

I. The question of the week

For two years the conversation about artificial intelligence and work has asked a single question with remarkable consistency: what will the machine do to us? Will it take the middle-class job, spare the poor one, lift the productivity number, close or widen the gap between the woman in the boardroom and the woman waiting to be let in? The question assumes a finished object — an intelligence that arrives, autonomous and complete, and then acts upon a labor market that receives it. What the question almost never asked, until very recently, is the prior one: who made the machine, and under what terms?

This is the arc worth tracing, because the two questions have been traveling on separate tracks and only lately collided. The labor that produces AI — the annotators labeling images for a few cents a task, the content moderators in Nairobi reviewing the worst of the internet so a chatbot can sound wholesome, the miners in Indonesia pulling tin from the ground — was documented early and thoroughly. It simply did not organize the discourse. The discourse belonged to readiness, upskilling, and the wealth engine.

Then, across a handful of quarters, the frame inverted. Where the record shows optimistic framing outrunning critical framing seven-to-one in early 2025, by the third quarter of 2026 the ratio had reversed and hardened, the word exploitation moving from footnote to headline. This piece follows that reversal: what we kept saying while the reality sat in plain sight, what was actually happening the whole time, and why the collision, when it finally came, arrived under the older and heavier vocabulary of colonialism rather than the softer language of disruption.

II. What we’ve been saying

The dominant register from late 2024 through the middle of 2025 was preparation. AI was coming; the task was to get ready for it. The World Economic Forum opened 2025 with the finding that only 2 percent of firms are prepared for large-scale AI adoption, framing the gap as a readiness challenge to be met with “talent transformation” and “responsible AI practices.” The verb doing the work there is transform — the worker as raw material to be upgraded, the firm as the site of adaptation. Nothing in that vocabulary points backward toward the people who built the system; it points only forward, toward the people who must now catch up to it.

The most striking feature of this stretch is how the labor question was consistently posed as a question of effect rather than production. Brookings asked whether a technology transforming middle-class jobs could also help the poor; the ABC asked what Australia could do to ensure workers benefit equally; the Harvard Business Review turned the anxiety inward, arguing that employees won’t trust AI if they don’t trust their leaders. Each is a serious piece. Each treats AI as a weather system — something that arrives, distributes its benefits and harms unevenly, and must be managed by the wise. None asks who is standing in the rain to make the weather.

By the second quarter of 2025 the optimism reached its rhetorical peak. Google put a number on the promise: £400 billion of economic growth available to the UK from AI-powered innovation, half of it contingent on workers actually adopting the tools. RealClearMarkets proposed AI as a wealth engine for women. TechBullion invited readers to embrace the AI revolution. The genre was aspirational, and the aspiration was individual: literacy, adoption, participation. Even the more sober contributions — the growing insistence that AI literacy is becoming a must — located the burden on the person facing the tool, not on the arrangement that produced it.

There was, to be fair, a corrective murmur inside the mainstream itself. The International Labour Organization, in a piece dated December 2024, described AI’s success as hinging on an invisible workforce performing low-paid, precarious tasks under challenging conditions — the word illusion sitting right in the title. This is worth pausing on. The critical account was not absent at the beginning; it was present, published by one of the most authoritative labor bodies in the world, and it did not move the center of gravity. It was catalogued alongside the optimism and outweighed by it. The conversation had the facts and declined to be reorganized by them.

Our own coverage registered the pattern without yet naming its cause. An earlier essay in this publication’s social-aspects series, the critical analysis of 4 May 2025, noted that authors writing on AI and equity tended to emphasize “transformative potential” — healthcare access, educational leveling — with the harms treated as a manageable second clause. By the analysis of 29 June 2025, the same series was tracking a rhetoric organized around “promoting equity, enhancing access to resources, and mitigating biases,” a vocabulary in which exploitation of the workforce that trains the models simply had no natural place to sit. The frame could hold bias — a defect in the output — far more comfortably than it could hold labor — a defect in the arrangement.

By late 2025 a subtler move appeared: the “human-centred” turn. Deloitte urged organizations to shift focus from technological prowess to the human element. It sounded like the correction the arc needed. But the human at the center of the human-centred approach was the user, the adopter, the employee whose buy-in determined whether the £400 billion materialized — never the annotator in Nairobi. The vocabulary of the human was recruited to serve adoption, not to indict production.

III. What’s been happening

The reality the rhetoric kept missing was neither hidden nor new. It had been mapped, in detail, before the current wave of optimism began. Kate Crawford’s account remains the clearest: the labor force of AI, she writes in The Atlas of AI (2021), runs “from the miners extracting tin in Indonesia to crowdworkers in India completing tasks on Amazon Mechanical Turk to iPhone factory workers at Foxconn in China,” a workforce “far greater” than the machine-learning engineers who receive the credit and the salaries. The extraction is the point. “Exploitative forms of work,” she writes, “exist at all stages of the AI pipeline, from the mining sector, where resources are extracted and transported to create the core infrastructure of AI systems, to the software side, where distributed workforces are paid pennies per microtask.”

Pennies per microtask. That phrase describes a stable material fact that persisted, unchanged, straight through the years of readiness rhetoric. While the WEF counted the 2 percent of firms prepared for adoption, the distributed workforce was already there, already paid, already invisible — not because it was concealed but because the story of autonomous intelligence has no role for it. Crawford’s insistence that labor is a form of extraction, continuous with the mining, is the analytic hinge: the human worker and the raw mineral occupy the same position in the supply chain, both fed into a system that presents its output as though it emerged from nowhere.

The institutions closest to labor saw this clearly and said so, even as their own future-of-work materials leaned optimistic. The ILO’s foundational framing distinguishes two applications of AI in the workplace — automating tasks workers perform, and deploying analytics over workers — but its December 2024 “illusion” piece went further, unpacking the precarity of the workers who make automation possible in the first place. The UN and ILO jointly warned of an AI divide in which the technology’s uneven adoption would track and deepen existing global inequality. The distributional facts were on the table by 2025: the American Enterprise Institute’s review of the emerging evidence found that 77 percent of adults in the highest-earning households had used AI recently — a use pattern tracking income and education so closely that the “democratizing” promise looked, on the numbers, like its opposite.

What changed was not the reality but the willingness to organize the conversation around it. Through 2025 the critical account remained a minority position even as evidence accumulated; the probe record shows optimistic framing outrunning critical framing by roughly seven to one in the first quarter, still better than two to one at the height of the boom in the second. The Global South contributions began to press harder — a 2025 framework on ethical and scalable AI for the Sustainable Development Goals written from the Global South, and a technical accounting of biases in AI and their implications for public administration that treated the harms as structural rather than incidental. Even MIT’s development-policy work, surveying how developing economies pursue AI strategy under resource constraints, framed the terms of participation as themselves unequal.

Then the frame broke open. By the third quarter of 2026 the critical count outnumbered the optimistic by more than four to one, and the vocabulary had escalated. The exposés no longer said precarious; they said exploitation, colonialism, consent. One account described the intelligence of modern AI as reliant on a global, outsourced “shadow workforce” and the uncompensated extraction of personal data — three intersecting crises, its authors argued, not three separate problems. Another, written by a student who remembered marveling at ChatGPT in high school, walked its readers back from the wonder to the unjust labor practices of the AI industry, naming the underpaid workers behind the fluent output. The document Crawford wrote in 2021 had, by 2026, become the frame the wider conversation finally adopted — five years late, and under duress.

IV. Where they meet, where they miss

The two tracks meet at the word illusion, and it is worth noticing that the ILO reached for it in December 2024, at the very start of the arc, while the wider discourse did not adopt it until late 2026. The gap between those two dates is the story. The facts did not need to be discovered; they needed to be believed, and belief lagged evidence by nearly two years. This is not a failure of information. It is a failure of frame — the readiness vocabulary could absorb almost any datum about who benefits from AI while remaining structurally incapable of holding the datum about who makes it.

The miss has a precise shape. The optimistic conversation was overwhelmingly about AI’s effect on labor: jobs displaced, jobs created, wages pressured, skills demanded. The critical reality was about labor as input to AI: the microtask, the annotation, the moderation shift, the mined tin. These are not two views of one thing. They are two different things, and the discourse spent two years discussing the first while the second sat undiscussed in the same room. When Deloitte urged a pivot from prowess to the human element, it named the right noun and pointed it at the wrong human — the adopter rather than the annotator. The correction that looked like arrival was a detour.

This publication has argued before that equity framing tends to convert structural questions into questions of access. The social-aspects analysis of 31 August 2025 noted the recurring claim that AI can “democratize access to education” and thereby level a playing field — a claim that treats the inequality as an information gap to be closed rather than a labor arrangement to be confronted. The hidden-labor arc is the sharpest test of that tendency. You cannot upskill your way out of being paid pennies per microtask; the problem is not that the worker lacks access to the technology but that the worker is the technology’s uncredited input.

Where the tracks genuinely converge — and this is the hopeful reading — is in the 2026 insistence that exploitation, colonialism, and data extraction are one crisis rather than three. That is Crawford’s argument arriving in the wild: labor as extraction, continuous with mining, all of it feeding a system that markets itself as autonomous precisely so it need not account for its inputs. The mystification is not accidental. Calling the product “intelligence” and the process “automation” performs specific work: it disappears the workforce. The 2026 exposés succeed to the degree that they refuse the vocabulary — naming the shadow workforce, insisting on consent, restoring the human whom “autonomous” was designed to erase.

The miss that remains is geographic and moral at once. The optimism was authored, overwhelmingly, from the places that consume AI; the exploitation is borne, overwhelmingly, by the places that produce its underlying labor. The AI divide the UN warned of is not only a gap in adoption. It is a gap between who narrates the technology and who is narrated by it — and for two years the narration ran entirely in one direction.

V. The longer view

The reversal in the record is real and it is welcome, but it should be read for what it is: the discourse catching up to a fact that its own most authoritative institutions had published at the outset. The correction did not come from new evidence. It came from a slow, grudging willingness to let evidence reorganize the story — and that willingness arrived only after the optimism had spent itself, after the wealth engine and the readiness gap had been talked through to exhaustion. There is a lesson in the sequencing. The critical account was available the entire time; what was scarce was not information but the frame that could bear its weight.

The task now is to keep the newer vocabulary from softening the way the old one did — to hold exploitation where the market would prefer challenge, and colonialism where it would prefer global supply chain. Crawford’s phrase resists that softening because it refuses abstraction: not a labor market, not a divide, but a person completing a task for a coin. The moment the conversation returns to weather — to the machine as a thing that simply arrives — the workforce disappears again.

So the sentence to carry home is the one the arc took two years to earn: the machine is not autonomous, it is populated, and every time we call it intelligent we are choosing not to count the people paid pennies to make it seem so.

References

1. Artificial intelligence — International Labour Organization.
2. Who Benefits From AI? New Studies Offer an Answer — American Enterprise Institute.
3. Mind the AI Divide: Shaping a Global Perspective on the Future of Work — United Nations and International Labour Organization.
4. The Artificial Intelligence illusion: How invisible workers fuel the “automated” economy — International Labour Organization (17 December 2024).
5. Global AI and the future of work — Forbes India (17 December 2024).
6. Unlocking human potential: Building a responsible AI-ready workforce for the future — World Economic Forum (20 January 2025).
7. Artificial intelligence is transforming middle-class jobs. Can it also help the poor? — Brookings (20 January 2025).
8. The AI revolution is coming — what can Australia do to ensure workers benefit equally? — ABC Religion & Ethics (23 March 2025).
9. Employees Won’t Trust AI If They Don’t Trust Their Leaders — Harvard Business Review (30 March 2025).
10. AI’s potential to tackle the UK’s productivity puzzle — Google (27 April 2025).
11. AI Has the Potential To Be a Wealth Engine for Women — RealClearMarkets (16 June 2025).
12. Why AI Literacy Is Becoming a Must — ResearchGate (29 June 2025).
13. Embracing the AI Revolution: Innovations Transforming the Future of Work — TechBullion (11 May 2025).
14. Are we ready to meet the expectations of AI for development? — Brookings (27 July 2025).
15. Unlocking AI’s Potential: A Human-Centred Approach — Deloitte (28 September 2025).
16. Determinants of Ethical and Scalable AI for the Sustainable Development Goals: A Qualitative Framework from the Global South — SPAST (20 July 2025).
17. Biases in Artificial Intelligence and Implications for AI Use in Public Administration: A Technical Perspective — Authorea (5 October 2025).
18. AI Strategy in Developing Economies — Internet Policy Research Initiative — MIT (6 September 2026).
19. The Crisis of AI Ethics: Exploitation, Colonialism, and Consent — Halt the Harm (6 September 2026).
20. Economy and Exploitation: The AI Industry’s Unjust Labor Practices — UAB Human Rights (30 August 2026).
21. The Atlas of AI: Power, Politics, and the Planetary Costs — Kate Crawford (2021).