AI and Its Invisible Workers
- Editorial Staff
- 11. 8.
- Minut čtení: 13
Aktualizováno: 21. 8.
Extraction, power, and human-fueled labor in menial, repetitive tasks
Prepared August 2025
Author: Strategic Analysis Unit – CEPRODE EUROPE
Download the full report here:
Executive Summary
This study examines the relationship between artificial intelligence (AI) and its workers, focusing on those who perform menial and repetitive tasks across the entire value chain: from the extraction of raw materials and hardware production, to logistics, and to the digital work of annotation, evaluation, and content moderation. Our core claim is that the real stakes of AI lie in the entanglement of global systems of extraction and power on which it relies and which it helps to reinforce.
Artificial intelligence has a people problem. Not because there are too few machine‑learning engineers—though they are scarce and richly paid—but because there are far more human hands than the industry likes to admit holding up the edifice. This study follows AI’s supply chains from the salt flats of the Andes to Chinese electronics plants and European and U.S. data centers, and into the browser windows of millions of digital workers who label, filter and correct the data that models ingest.
The thesis is blunt: behind the sleek apps and headline‑grabbing models sits a web of extraction—of minerals, water, energy and human labor—that AI depends on and, in many cases, reinforces. ‘Automation’ is not the end of work so much as a redistribution of it, the authors argue, away from celebrated campuses and toward overlooked sites at the edge of global markets.
We show that AI is rooted in material processes with high energy and water intensity (e.g., data centers and cloud infrastructures) and in supply chains that involve lithium, cobalt, rare earths, and tin. We document already visible environmental impacts and future risks at extraction sites (e.g., Andean salt flats, mines in the Democratic Republic of the Congo, rare‑earth tailings basins in China) and at nodes of energy consumption (data centers powered by electricity mixes that in many regions are still largely fossil). AI’s ecological footprint is often undercounted or under‑communicated, and environmental costs are rarely internalized.
On the labor side, AI is not only highly paid engineers. It is also miners in Indonesia and in the Democratic Republic of the Congo, workers in Chinese electronics factories, warehouse operatives in logistics nodes, and a vast army of digital workers (crowdworkers) and moderators who label, correct, and “feed” AI systems. The “brains” of AI are trained on data labeled by crowdworkers: ImageNet, one of the most influential datasets, was built via Amazon Mechanical Turk (AMT). We show how contract and outsourced labor often outnumbers direct employees in large technology firms, with fewer benefits and scant job security. We also explore the “automation illusion,” which depends on thousands of people in the “last mile” to make systems appear interruption‑proof. The illusion of automation rests on thousands who label, correct, evaluate, and “patch” the systems: the “artificial artificial intelligence” popularized by Jeff Bezos to describe MTurk.
Finally, we offer an economic estimate—expressed as a range—of the sector’s excess profits attributable to low wages and the exploitation of unpaid labor (e.g., reCAPTCHA puzzles and underpaid micro‑tasks), discussing assumptions, sources, and methodological limits. The (conservative) economic estimate of excess profits is: (a) for paid micro‑work, using ILO data (global average remuneration of $3.31/hour including idle time), the shortfall relative to an “ethical floor” of $6.50/hour yields a gap of $3.19/hour; multiplied by millions of hours per year, this produces billion‑scale transfers of value from labor to profits; (b) unpaid work through reCAPTCHA may be worth ~USD 1.5–3 billion per year (assumption: 200 million challenges/day, ~10 seconds each; valuation between the U.S. minimum wage and $15/hour).
We conclude with policy and corporate recommendations that recognize and remunerate the human labor sustaining AI, increase transparency and accountability across material and digital supply chains, and internalize ecological costs.
1. Introduction: power, extraction, and AI
This study examines the relationship between AI and its workers with an exclusive focus on the humblest and most repetitive tasks. We adopt an extractivist perspective: AI as a technopolitical complex that connects the extraction of resources (raw materials, energy, water), the extraction of data, and the extraction of labor (paid little or not at all).
Recent scholarship in the political ecology of AI insists that AI systems are not “immaterial”: they depend on resource extraction, global production chains, energy‑hungry infrastructures, and vast amounts of human labor. Within this frame, we focus selectively on the “invisible” workers who perform menial and repetitive tasks—often precarious and poorly paid—without which many AI services would not function. The document is organized as follows: (i) the material supply chain of AI; (ii) human labor across hardware, logistics, and platforms; (iii) “human‑fueled automation”; (iv) an exploratory quantitative analysis of excess profits; and (v) conclusions and recommendations.
2. Methodology and definitions
The study adopts a qualitative–quantitative approach: a synthesis of evidence from institutional reports, investigative journalism, and academic literature; plus an estimation exercise (scenarios) of the economic value of underpayment and unpaid labor. By “menial and repetitive jobs” we mean activities of extraction, assembly, handling, labeling, moderation, and short‑duration micro‑tasks with limited autonomy. Quantitative results should be interpreted as plausible ranges subject to uncertainty; we provide all assumptions.
Crawford and Joler, in “Anatomy of an AI System” (the Amazon Echo case), map the device as a node within a global network that cuts across mines, assembly lines, server farms, crowdwork platforms, and massive data collection. The diagram makes visible what the interface hides: energies, materials, and bodies.
3. The material extraction of AI and environmental harms
AI rests on supply chains for critical minerals and on infrastructures with high energy and water intensity. Key materials include lithium (Andean brines, including Bolivia), cobalt (the Democratic Republic of the Congo), rare earths (extraction and refining, with significant externalities in areas such as Baotou, China), and tin (Indonesia’s Bangka‑Belitung islands, used for electronic solder). Downstream, data centers and networks consume electricity and water for power and cooling; the footprint depends on the local energy mix and on efficiency practices.
Lithium (Bolivian Andes and the Lithium Triangle)
Lithium—critical for the batteries and electronics that power AI—largely comes from the Lithium Triangle (Bolivia, Chile, Argentina). Scientific studies document high water consumption and risks to ecosystems, particularly in the Andean salt flats, with effects on Indigenous communities and fauna (e.g., declines in flamingo populations). In Bolivia, Indigenous organizations already denounce conflicts over water and extraction processes carried out without adequate social and environmental safeguards.
Key points (lithium):
Evaporation of brines and freshwater use; estimates of tens of m³ per tonne of lithium carbonate at sites such as the Salar de Atacama (Chile).
Bolivian communities (Uyuni and surroundings) report pressure on water resources and limited consultation.
The “green” rhetoric rarely accounts for long‑term costs for aquifers, biodiversity, and livelihoods.
Cobalt (DRC)
Substantial evidence of child labor, hazards, and pollution accompanies artisanal cobalt extraction in the Congo, a key component of lithium‑ion batteries. Amnesty International (2016) and subsequent institutional reports document health risks, accidents, and opaque supply chains that feed global electronics.
Tin (Indonesia) and rare earths
Tin (for electronic solder) extracted in the Bangka‑Belitung islands has left “lunar” landscapes, destruction of coral reefs, and conflicts with fisheries; the move toward offshore extraction exacerbates damage to marine ecosystems.
Rare earths and other critical metals share significant environmental externalities in extraction and refining stages; these costs are seldom internalized by those who monetize final products.
Table 1. Critical materials and flows for AI: areas and impacts (summary)
The material footprint of AI: energy, water, and data centers
The explosive growth of AI is pushing the electricity consumption of data centers upward: the IEA estimates that data‑center electricity demand will roughly double by 2030, with AI uses as a principal driver; figures on the order of ~945 TWh in 2030 are under discussion (more than today’s electricity consumption of Japan).
In parallel, journalistic and technical analyses point to a growing dependence on natural gas to support new compute capacity, contradicting “100% renewables” narratives based on credits that are not temporally aligned with real‑time consumption.
Water. AI also has a water footprint (cooling) that is often underestimated in environmental balance sheets. (For the electricity dimension alone, see IEA estimates and methodological debates on how to project AI‑specific consumption.)
4. Human labor along the value chain: mines, factories, logistics, platforms
Alongside extraction, hardware assembly takes place in large plants (e.g., supply chains linked to Foxconn for smartphones), while logistics and e‑commerce make extensive use of “hybrid” human–robot labor with stringent algorithmic metrics (e.g., Amazon: Time off Task, or TOT, and ambiguity about other timing metrics). In Amazon warehouses, TOT—Time off Task—is documented as a metric to monitor and discipline workers, with penalties at times based on incorrect system assumptions. “Time to Live” (TTL), by contrast, is a technical term in networking and storage lifecycles and does not describe worker surveillance. In labor governance, the relevant concept is TOT, not TTL.
In the “digital” domain, millions of crowdworkers and moderators classify images, texts, emotions, identities, and gestures; perform quality checks; rename and clean data; evaluate model responses (RLHF); and moderate harmful content. This workforce is often contracted, outsourced, without benefits, and with scant protection.
5. Examples and cases
“Artificial Artificial Intelligence”: the historic motto of Amazon Mechanical Turk (AMT) indicating that so‑called artificial intelligence is often sustained by low‑cost human intelligence.
Micro‑tasks and moderation: investigations and ILO studies estimate low effective pay (often a few dollars per hour, net of unpaid time) for annotation and verification tasks on platforms such as AMT, Figure Eight/Appen, Microworkers, and Clickworker.
Pseudo‑automation: cases such as x.ai (“Amy/Andrew”) and other assistants reveal the use of human labor to cover the limits of conversational AI.
Logistics: the adoption of robots (e.g., Sparrow, Proteus) does not eliminate human labor in warehouses; rather, it reassigns tasks and introduces new speed and control constraints.
reCAPTCHA: the puzzles users solve to “prove they are not robots” have also contributed to the creation/improvement of datasets for address reading and image recognition (e.g., Street View), constituting a case of diffuse, unpaid labor.
6. Human‑fueled automation and the “automation illusion”
A growing body of research describes the “last mile” of automation as dependent on pervasive human interventions: from labeling to the continuous evaluation of models, to real‑time quality control. Behind the appearance of autonomy lies a multitude of micro‑contributions that keep the system stable and “presentable.” This holds in warehouses (exception handling), in contact centers (human escalations), on annotation platforms, and in content‑moderation services.
Logistics nodes: the illusion of automation. In warehouses and supply chains, humans close the loop on what machines do not do “well enough”: selecting irregular objects, handling exceptions, and corrections. Technological supply chains remain human‑supervised despite the rhetoric of total automation. (See also the “Anatomy of an AI System” map.)
7. Employment structures and asymmetries (contractors vs employees)
Large technology companies rely extensively on contract personnel (TVCs, outsourcers), often more numerous than direct employees. This entails gaps in wages, benefits, and protections. On platforms, contractual fragmentation reduces both individual and collective bargaining power, increasing income volatility and shifting operational risk onto workers.
What crowdworkers do and how much they earn
Many foundational AI datasets are built from crowdsourced annotations: ImageNet, for instance, was collected/validated via Amazon Mechanical Turk.
UN/ILO surveys (3,500 workers in 75 countries; platforms: AMT, CrowdFlower/Figure Eight, Microworkers, Clickworker, Prolific) show very low hourly pay and significant unpaid time (task search, waiting, rework). In 2017 the global average was $4.43/hour for paid time only and $3.31/hour when unpaid hours are included; the median falls to $2.16/hour considering paid + unpaid time.
Jeff Bezos labeled MTurk “Artificial Artificial Intelligence,” acknowledging that the “intelligence” is, in reality, supplied by humans behind the interface.
Why is the dependence on underpaid labor ignored?
Invisibilization (interfaces erase the human presence);
Meritocratic/technological narratives that ascribe value solely to the algorithm;
Fragmentation of labor into non‑negotiated micro‑tasks, which lowers bargaining power (the ILO also documents non‑payment and scarce avenues for redress).
Unpaid labor: the reCAPTCHA case and other examples
reCAPTCHA (acquired by Google) “recycled” the human verification gesture to train OCR, Street View, and, more generally, vision systems: for years roughly 200 million CAPTCHAs were solved per day, at ~10 seconds each. It is unpaid labor, yet with economic value.
Other services have at times masked human labor as “AI” (scheduling assistants and “Wizard‑of‑Oz” arrangements): the x.ai (Amy/Andrew) case remains emblematic of the blurred boundary between automation and manual oversight.
8. Economic estimate of excess profits from low wages and unpaid labor
We propose a scenario‑based estimate, separating (A) annotation/evaluation micro‑tasks, (B) content moderation, and (C) diffuse unpaid activities such as reCAPTCHA puzzles. The principle is to compare observed remuneration with a share of “fair compensation” (a benchmark), or to assign a monetary value to time spent without compensation. The results are plausible ranges that indicate orders of magnitude.
Main assumptions (summary):
• Online Labour Observatory (OII) estimates ~14 million active workers in online labor (all platforms). We assume 25% perform AI‑related micro‑tasks at least occasionally.
• Annual per‑capita hours on micro‑tasks: low scenario 250; medium 500; high 750.
• Effective average wage for micro‑tasks: $3/hour; “fair” benchmark: $6/hour (double, to include unpaid time and minimum protections).
• Moderation: global orders of magnitude ranging from tens of thousands to a few hundred thousand staff. We assume 50k (low), 100k (medium), 200k (high); 2,000 annual hours; effective average wage $3/hour; benchmark $15/hour.
For unpaid labor (reCAPTCHA):
• Volume: 50/100/200 million challenges/day × 10 seconds = ~555,000 hours/day ≈ 50.75/101.5/203 million hours/year.
• Valuation: from $7.25/hour (U.S. federal minimum) to $15/hour; applied average $11.12/hour.
• Order of magnitude: $0.56–2.25 billion/year.
Table 2. Scenario estimate of annual excess profits attributable to low wages and unpaid labor
9. Discussion: limits, sensitivities, and implications
These estimates aggregate heterogeneous markets with partial information about headcounts and geographic composition. Assumptions about shares and per‑capita hours materially affect results; hence, three scenarios are provided. The reCAPTCHA figure is sensitive to uncertainty over the current volume of puzzles and over the fraction that is effectively “useful” for dataset creation. Overall, the orders of magnitude (billions of USD per year) are robust to reasonable parameter variations.
Joint reading (prudent range): between $2.8 and $19.6 billion per year when summing (i) a minimum proxy for unpaid labor via reCAPTCHA and (ii) the wage gap over a plausible range of global micro‑labor hours. The order of magnitude is in the billions even without accounting for environmental damages, health externalities, or proprietary labeling programs. (For data centers, the operational continuity and extra energy needed for AI are rising fast, amplifying further excess profits derived from economies of scale.)
Methodological caveats. The estimates are conservative and intended to identify the scale of value transfer: there is no exhaustive public accounting of hours and rates for all industrial actors; the calculations do not include external costs (water, biodiversity, health). Where possible, we have adopted minimal assumptions and primary sources (ILO; public/technical statements on reCAPTCHA and IEA for consumption).
Note: these estimates do not include (a) environmental externalities (lithium, cobalt, tin), (b) healthcare/psychosocial costs of content moderation (e.g., Kenya cases), or (c) additional energy and fixed capital for data centers. The overall “social cost” is therefore higher.
10. Conclusions and recommendations
Contemporary AI rests on a combination of material extraction, energy‑hungry infrastructure, and diffuse human labor.
The real stakes of AI are political and economic: an entanglement of global systems of extraction and power. Innovation relies on material and human supply lines, and produces externalities that are not adequately accounted for (energy, water, biodiversity; occupational health and safety).
AI work ≠ only engineers. From the mines of the Congo and the Indonesian islands to assembly labs and warehouses, and to Indian and American annotators of datasets, AI mobilizes a mosaic of menial and repetitive jobs. Without them, the systems would not exist.
Human‑fueled automation. Behind the “models” there are always humans who label, correct, evaluate, and moderate. This dependence remains culturally underestimated and institutionally under‑regulated (non‑payment, limited remedies, low transparency).
Excess profits: wage compression and the appropriation of unpaid labor generate billions of dollars per year transferred from labor to profits, even under prudential scenarios. The reCAPTCHA case alone suggests orders of magnitude > $2.2 billion/year; multiplying the wage gap by hundreds of millions of micro‑labor hours brings the overall sum into the multi‑billion range.
For workers performing menial and repetitive tasks—physical and digital—the gap between generated and recognized value remains wide. To close it, we suggest: (i) mandatory transparency on material supply chains and on the human‑labor content of AI systems (including metrics on hours, wages, and geographies); (ii) minimum standards of compensation and protections for crowdwork and moderation, with joint liability between principals and intermediaries; (iii) internalization of environmental costs through explicit pricing (carbon, water) and local compensation funds; (iv) independent audits of the “last mile” of automation, with recognition of human contributions; (v) investments in energy efficiency, scheduling on low‑emissions grids, and public reporting of training and inference footprints.
If we take seriously the maxim that “AI is only as good as its data,” we must add: …and as good as the human labor that produces them. Making this supply chain visible—from salt flats to GPUs, from micro‑tasks to warehouse aisles—is the first step toward AI that pays fairly, compensates harm, and redistributes the benefits of automation beyond the glittering interface.
What would success look like? Not a return to artisanal computing, but an industry that acknowledges its human and environmental foundations and prices them honestly. Models would still get built; progress would continue. But the people who mine, solder, ship, label and moderate would show up in the story—and on the balance sheet. The study’s provocation is simple: AI will be judged not only by what it can do, but by what it chooses to pay for.
Appendix A. Numerical assumptions and calculations
• Active online workers (all platforms): 14 million (OII estimate). Micro‑task share (10–40%).
• Hours/year per capita for micro‑tasks: 250–750. Effective wage: $3/h; benchmark: $6/h.
• Moderators: 50k–200k globally; 2,000 hours/year; effective wage $3/h; benchmark $15/h.
• reCAPTCHA: 50–200 million puzzles/day; 10 s/puzzle; value of time from $7.25/hour (U.S. minimum) to $15/hour; applied average $11.12/hour.
The resulting total range spans, illustratively, from a few billions to over nineteen billion dollars per year.
Essential bibliography (selected)
Crawford, K. & Joler, V. (2018). Anatomy of an AI System. https://anatomyof.ai
ILO (2018). Digital labour platforms and the future of work: Global survey of 3,500 crowdworkers in 75 countries. https://www.ilo.org/wcmsp5/groups/public/---dgreports/---inst/documents/publication/wcms_645337.pdf
IEA (2024). Data centres and AI. https://www.iea.org/reports/data-centres-and-data-transmission-networks
Nature (2021). Rare earth mining impacts (Tailings, Baotou). https://www.nature.com/articles/s41893-021-00756-9
Amnesty International (2016). ‘This is what we die for’ (cobalt in DRC). https://www.amnesty.org/en/documents/afr62/3183/2016/en/
Mongabay (2025). Communities vs lithium in Bolivia. https://news.mongabay.com/2025/04/bolivian-communities-push-back-against-foreign-backed-lithium-projects/
Google Security Blog (2014). Street View & reCAPTCHA. https://security.googleblog.com/2014/04/hard-to-read-recaptcha-phrases-now-using.html
Bloomberg (2016). The humans hiding behind the chatbots (x.ai). https://www.bloomberg.com/news/articles/2016-04-18/the-humans-hiding-behind-the-chatbots
Vox (2019). Google’s white- and red-badges (contractors). https://www.vox.com/recode/2019/5/28/18642805/google-contractors-employees-headcount
Business Insider (2025). Amazon robots & humans. https://www.businessinsider.com/how-amazon-uses-robots-sort-transport-packages-warehouses-2025-2
Amazon (2024). Robotics overview (Sparrow, Proteus). https://www.aboutamazon.com/news/operations/amazon-robotics-robots-fulfillment-center
OII – Online Labour Observatory (2021/2023). Estimates on active online workers. https://ilabour.oii.ox.ac.uk/how-many-online-workers/
Appen (2023). Annual Report (FY2023). https://www.annualreports.com/HostedData/AnnualReportArchive/a/ASX_APX_2023.pdf
Bloomberg (2025). Scale AI revenue (2024 estimate). https://www.bloomberg.com/news/articles/2025-04-02/scale-ai-expects-to-more-than-double-sales-to-2-billion-in-2025
The Guardian / BBC / Friends of the Earth (2012–2014). Tin in Indonesia (Bangka‑Belitung). https://www.theguardian.com/environment/2012/nov/23/tin-mining-indonesia-bangka
Appendix B. Amazon: 'Time to Live' (TTL) and 'Time off Task' (TOT)
The term “Time to Live” (TTL) in Amazon/AWS denotes a technical characteristic of data lifecycles: in services such as DynamoDB or SageMaker’s Feature Store, records can be marked with an expiration time to be automatically deleted beyond a given date/time. In cloud architectures, TTL is a governance tool for storage and costs, not a criterion for managing human labor.
By contrast, “Time off Task” (TOT) in Amazon’s logistics contexts is an algorithmic metric that quantifies the time during which a worker is not performing operations recorded by the system (scanners, stations). Organizational policies linked to TOT have affected work tempo and protections in warehouses. In this study we analyze TOT as an example of data‑driven labor governance and automation, whereas TTL remains a concept of information lifecycle management in AWS.
Methodological note: a study by the International Labour Organization (a UN agency) covering 3,500 crowdworkers from 75 countries (AMT, Figure Eight/Appen, Microworkers, and Clickworker) documents low effective remuneration and significant unpaid time, confirming the structural link between AI and micro‑task labor.
Komentáře