01 · The hidden workforce
Safety filters and training data are built by people, sometimes through outsourcing firms in lower-income countries. In one documented contract, journalists reported that workers’ take-home pay was a small fraction of what the AI company paid for their work; the firm disputed the figures. Ask who is in the chain and what each link receives.
source-assertionConfidence: mediumContestability: medium
An investigation found that Kenyan workers labelling harmful content for a chatbot’s safety filter took home roughly $1.32–$2 an hour while the contract paid the outsourcing firm $12.50 an hour; the firm disputed the take-home figures.
Our reading: One contract in one country, reported by journalists; the firm’s own figure was $1.46–$3.74 after tax. It shows how value is split along one chain, not the pay of all data workers.
Inspect evidence · 1 passage
Passage summaries are paraphrases, not quotations. Open the source to read the full passage.
Exclusive: OpenAI Used Kenyan Workers on Less Than $2 Per Hour to Make ChatGPT Less Toxic (opens in a new tab)Pay and contract paragraphs
Kenyan workers labelling harmful text for a chatbot safety filter through an outsourcing firm took home roughly $1.32–$2 an hour, while the contract rate paid to the firm was $12.50 an hour; the firm said workers could earn $1.46–$3.74 after tax. Workers described the work as psychologically taxing.
Scope note: Investigative journalism. The outsourcing firm disputed the pay figures; both ranges are recorded.
Passage checked 2026-10-18 · time-kenya-pay
03 · Whose view gets labelled into the data
Instructions tell workers how to label. When work is precarious, workers follow instructions rather than question them, so the client’s view of the world is built into the data. Working conditions and data quality are linked.
source-assertionConfidence: mediumContestability: medium
Research with data workers in Latin America found that task instructions impose clients’ views on how data is labelled, and that precarious conditions make workers follow them rather than question them.
Our reading: Qualitative study of specific platforms and one outsourcing firm; it explains a mechanism rather than measuring its prevalence.
Inspect evidence · 1 passage
Passage summaries are paraphrases, not quotations. Open the source to read the full passage.
The Data-Production Dispositif (opens in a new tab)Abstract
Analysing 210 instruction documents and 55 interviews, the study finds that task instructions impose clients’ worldviews on data, and that precarious work and economic dependency make workers obedient to those instructions.
Scope note: Qualitative study of data workers in Venezuela and Argentina.
Passage checked 2026-10-18 · miceli-posada-abstract
interpretationConfidence: mediumContestability: medium
Data work is part of AI infrastructure: the pay, protection and voice of the people who label and moderate shape both who bears AI’s costs and the quality of the systems built on their work.
Our reading: AIGovSpec’s interpretation, placing labour inside the infrastructure object and the distribution test.
Inspect evidence · 3 passages
Passage summaries are paraphrases, not quotations. Open the source to read the full passage.
Exclusive: OpenAI Used Kenyan Workers on Less Than $2 Per Hour to Make ChatGPT Less Toxic (opens in a new tab)Pay and contract paragraphs
Kenyan workers labelling harmful text for a chatbot safety filter through an outsourcing firm took home roughly $1.32–$2 an hour, while the contract rate paid to the firm was $12.50 an hour; the firm said workers could earn $1.46–$3.74 after tax. Workers described the work as psychologically taxing.
Scope note: Investigative journalism. The outsourcing firm disputed the pay figures; both ranges are recorded.
Passage checked 2026-10-18 · time-kenya-pay
Fairwork Cloudwork Ratings 2025 (opens in a new tab)pp. 3–4 and conclusion (p. 34)
Of 16 online labour platforms rated in 2024–25, evidence of policies ensuring the local minimum wage was found for only four; the authors conclude the platforms remain far from safeguarding basic fair-work standards and that unfair and insecure work is the norm on most.
Scope note: Research ratings of 16 online labour platforms against five fair-work principles; fieldwork February 2024 to January 2025. The final part of the report was not read.
Passage checked 2026-10-18 · fairwork-cloudwork-findings
The Data-Production Dispositif (opens in a new tab)Abstract
Analysing 210 instruction documents and 55 interviews, the study finds that task instructions impose clients’ worldviews on data, and that precarious work and economic dependency make workers obedient to those instructions.
Scope note: Qualitative study of data workers in Venezuela and Argentina.
Passage checked 2026-10-18 · miceli-posada-abstract
04 · Exposure is not job loss
Research on jobs measures how many tasks generative AI could affect, not how many jobs will go. Exposure is spread unevenly across countries and between women and men. Distribution depends on choices employers, workers and governments make next.
source-assertionConfidence: highContestability: medium
Generative AI exposure falls unevenly: women’s jobs and jobs in richer countries are more exposed, and change within jobs is more likely than outright replacement.
Inspect evidence · 1 passage
Passage summaries are paraphrases, not quotations. Open the source to read the full passage.
Generative AI and Jobs: A Refined Global Index of Occupational Exposure (opens in a new tab)Abstract and key findings
About one in four workers worldwide is in an occupation with some generative AI exposure. The highest-exposure category covers 4.7% of women’s jobs against 2.4% of men’s, and exposure rises with country income. Transformation of jobs is more likely than replacement.
Scope note: Intergovernmental and tripartite (governments, employers, workers). Measures exposure, not realised job loss.
Passage checked 2026-10-09 · ilo-exposure-uneven
source-assertionConfidence: highContestability: low
Labour organisations propose making worker voice a legal requirement: compulsory union consultation before high-risk workplace AI is deployed, plus a right to human review.
Inspect evidence · 1 passage
Passage summaries are paraphrases, not quotations. Open the source to read the full passage.
The AI Bill Project (opens in a new tab)Summary of the draft Artificial Intelligence (Regulation and Employment Rights) Bill
The draft bill would give unions a statutory right to be consulted before employers deploy high-risk AI decision systems, and give workers a right to human review and a personal explanation of AI decisions.
Scope note: Labour voice. Model legislation drafted by a union federation; it is a proposal, not law.
Passage checked 2026-10-09 · tuc-consultation-right
05 · Different tests, different blind spots
These lenses overlap; they are not mutually exclusive schools. We have named them to expose what each reading brings into view.
Workers as investigators
Data workers document their own conditions and set the research questions, publishing their findings rather than being studied by others.
Our reading: A labour-led project hosted by research institutes; it reports workers’ own accounts rather than representative statistics.
Inspect evidence · 1 passage
Passage summaries are paraphrases, not quotations. Open the source to read the full passage.
Data Workers’ Inquiry (opens in a new tab)About the project
Data workers in several countries set the questions and publish their own findings about their workplaces, as a participatory alternative to research done on them.
Scope note: Labour voice: participatory research in which data workers investigate and publish on their own workplaces. Launch date from the Weizenbaum Institute announcement.
Passage checked 2026-10-18 · data-workers-inquiry-method
Fair-work standards
Independent ratings hold platforms to minimum standards on pay, conditions, contracts, management and representation, and find most falling short.
Our reading: Ratings depend on platforms’ cooperation and the researchers’ evidence thresholds; a low score means evidence of a standard was not found.
Inspect evidence · 1 passage
Passage summaries are paraphrases, not quotations. Open the source to read the full passage.
Fairwork Cloudwork Ratings 2025 (opens in a new tab)pp. 3–4 and conclusion (p. 34)
Of 16 online labour platforms rated in 2024–25, evidence of policies ensuring the local minimum wage was found for only four; the authors conclude the platforms remain far from safeguarding basic fair-work standards and that unfair and insecure work is the norm on most.
Scope note: Research ratings of 16 online labour platforms against five fair-work principles; fieldwork February 2024 to January 2025. The final part of the report was not read.
Passage checked 2026-10-18 · fairwork-cloudwork-findings
Related context: exposure, not only loss
Labour-market research measures which jobs are exposed to generative AI and expects most to change rather than disappear, with exposure falling unevenly across countries and between women and men.
Our reading: This perspective concerns jobs changed by generative AI, not data work itself; it is included because the guide covers both. Exposure is not job loss.
Inspect evidence · 1 passage
Passage summaries are paraphrases, not quotations. Open the source to read the full passage.
Generative AI and Jobs: A Refined Global Index of Occupational Exposure (opens in a new tab)Abstract and key findings
About one in four workers worldwide is in an occupation with some generative AI exposure. The highest-exposure category covers 4.7% of women’s jobs against 2.4% of men’s, and exposure rises with country income. Transformation of jobs is more likely than replacement.
Scope note: Intergovernmental and tripartite (governments, employers, workers). Measures exposure, not realised job loss.
Passage checked 2026-10-09 · ilo-exposure-uneven
06 · Worked example · fictional
A small firm hires a labelling vendor
Fictional exercise. A small company building a customer-service assistant needs thousands of example conversations rated for quality and safety. The cheapest quote comes from a vendor that uses an online platform with workers in several countries. No real company, vendor, platform or country is assumed.
1. Who will do the work, and on what terms?
Ask the vendor who the workers are, how they are paid per hour (not per task), and whether they can refuse distressing content. A vendor that cannot answer is telling you something.
2. What does the price leave out?
The cheapest quote may leave out breaks, support for workers who review harmful content, or time to query unclear instructions. Those costs do not disappear; they move to the workers.
3. Whose judgment ends up in the data?
Write instructions that invite questions and record disagreements between raters. Treat workers’ feedback as information about your system, not as noise.
4. Which object would each fix act on?
Minimum pay terms in the contract act on incentives. An audit of the vendor or a certification is a safeguard. A worker voice mechanism or union recognition builds an institution. Sort your options this way to see what is missing.
5. Whose jobs change next?
The assistant may change the work of the company’s own support staff. Ask whether they were consulted and what happens to their roles.
Buying data work buys a set of working conditions. A buyer can change incentives through contract terms, demand safeguards such as audits, and support institutions that let workers be heard, or it can pass the costs down the chain without seeing them.
What this lesson does not settle
This draft lesson has not yet been reviewed by the framework’s author or an outside reviewer. Its evidence comes from one investigation, one ratings report, one qualitative study and one labour-market index; it does not measure data workers’ conditions worldwide. Court cases brought by content moderators in Kenya are not summarised here; their current status was not verified. Workers’ own accounts are represented through a participatory project, not through representative surveys.