AI Does Not Learn Alone: Why the ILO Is Pulling Latin American Data Work Into the Center of the AI Supply Chain
Original Chinese title: AI 看起來會自己學，其實有人半夜幫它「看懂資料」：ILO 把拉丁美洲 data work 拉到 AI 供應鏈正中央
ILO discussions of Latin American data work show that AI capability rests on a cross-border labour chain of annotation, review and payment; procurement and quality systems must not externalise equipment, scheduling and payment risks to workers.
鍾靜蓉
A National Taiwan University of Science and Technology doctorate in digital education focused on digital teaching strategies and metadata reasoning analysis.
The phrase “AI learns by itself” hides a production line
Model releases foreground parameters, benchmarks and new capabilities. They rarely show the people who select examples, label images, compare responses, remove harmful material, translate instructions or adjudicate ambiguous cases. The ILO’s INDL-9 discussion of AI supply chains, including Latin American data work, places that labour back inside the account of how artificial intelligence is produced.
The point is not that automation is unreal. It is that automated capability has a labour history. Human judgments are organised through platforms, subcontractors, payment systems and quality scores before they reappear as a seamless product. Ignoring that history distorts both the ethics and the engineering of AI.
Why supply chain is the right frame
A supply chain has upstream clients, intermediaries, production rules, quality inspection and uneven bargaining power. AI has all of these. Its inputs may be data and judgment rather than metal or grain, but work is still divided across firms and borders, and responsibility can still disappear through subcontracting.
Once AI is seen as a production system, familiar questions become unavoidable. Who sets the price? Who owns the equipment? Who bears the cost of rejected work? Can a worker appeal an automated quality decision? Does the lead company know how many intermediaries stand between its purchase order and the person completing the task?
Territory changes what a “microtask” costs
Remote work is often described as frictionless. In Latin America, the actual conditions may include unstable broadband, shared household computers, expensive electricity, currency conversion, account restrictions and schedules aligned with clients in another time zone. A task that looks small to a buyer can reorganise an evening for a worker and a family.
These costs are not accidental details outside the labour relationship. They are part of the infrastructure on which the client relies. When a platform pays only for accepted clicks, the worker may subsidise training, waiting, clarification, rejected output and equipment from an uncertain income.
Payment is part of the work, not an administrative afterthought
Julian Posada’s research on digital payments shows why the nominal task price can be misleading. Delays, exchange rates, withdrawal thresholds, fees and local financial access determine what reaches the worker and when. A wage is embedded in technical and financial systems.
An accountable platform should disclose gross price, deductions, conversion method, expected payment date and a usable dispute path. Workers need downloadable records for taxes and income verification. A company that audits annotation accuracy but cannot trace unexplained deductions is measuring only the side of quality that serves the buyer.
Better working conditions support better data
Data quality and labour conditions are not competing goals. If instructions are vague, questions cannot be asked and rejections are unexplained, workers must guess what the client intended. Inconsistent guesses become inconsistent labels, and those labels become model behaviour.
Versioned guidelines, representative boundary cases, paid training and a channel for clarification improve both fairness and reproducibility. When a rule changes, rework should be compensated. Human review should examine automated quality flags before they affect access to future tasks or earned pay.
Procurement must follow the work beyond the first contractor
Lead AI companies can require suppliers to disclose subcontracting layers, pay terms, exposure to harmful content, working-time expectations and grievance mechanisms. Contracts should permit worker interviews and evidence-based audits, not merely a signed code of conduct.
Traceability also improves data provenance. A buyer should know which guideline version governed a dataset, how disagreements were resolved, which languages were involved and what quality checks were applied. Without that history, the company cannot fully explain the origin of its training material.
What public policy needs to see
Data work often falls between familiar categories of employment and self-employment. Regulators should examine actual control and dependency: who assigns tasks, sets price, monitors performance, limits appeal and can remove access to income? The platform’s label should not decide the legal analysis in advance.
Basic visibility matters. Labour statistics, cross-border payment records and occupational-health discussions need categories capable of capturing data work. Harmful-content exposure, night work and unpredictable availability are health and safety issues even when the workplace is a kitchen table.
Education about AI should include the people who make data legible.
Students learning to use generative systems should also learn how classification rules are written and contested. Annotating an image is not merely pressing a button; it involves concepts, borderline cases, cultural context and feedback. That cognitive work becomes harder to see when a task is broken into tiny pieces.
This perspective connects technical literacy to labour literacy. It also helps explain why community, language and cultural data are not free raw material. Data can embody work, relationships and rights even when a file is easy to copy.
Conclusion: invisible work is still work
The ILO conversation gives public debate a more complete unit of analysis. Instead of asking only whether AI will replace jobs, it asks which jobs AI has already created, how those jobs are organised, and who carries the uncertainty that makes the product appear effortless.
Mature AI governance must connect procurement, payment, working conditions and model quality. A system cannot be called sustainable if its reliability depends on people whose labour remains deliberately difficult to trace.
Yuan Media AI | Continue by role
- Latin American data worker: Which unpaid costs and payment risks determine whether a task is viable?
- AI company data or procurement manager: How far down the contracting chain can conditions and provenance be audited?
- Labour-policy or ILO researcher: Which indicators make geographically distributed data work visible?
- Worker-rights organiser: How can isolated platform disputes become collective evidence without exposing workers to retaliation?
Sources
AI use and content-safety disclosure
This English edition is an AI-assisted translation of the supplied Chinese feature, checked for source parity and evidence boundaries.