{"course":{"slug":"ai-actually-explained","title":"AI, Actually Explained","url":"https://addaly.com/learn/ai-actually-explained","level":"beginner","track":"start"},"lesson":{"slug":"the-hands-that-trained-it","title":"The labour that trained it","url":"https://addaly.com/learn/ai-actually-explained/the-hands-that-trained-it","position":80,"of":92,"minutes":9,"answer":"The second training that makes a model helpful and safe is done by people ranking answers and labelling harm, much of it paid by the task in Kenya, the Philippines, Venezuela and India.","keyIdea":"The second training that makes a model helpful and safe is done by people ranking answers and labelling harm, much of it paid by the task in Kenya, the Philippines, Venezuela and India, and the price of a query hides that labour the way it hides the electricity."},"body":{"format":"markdown","text":"## Where the manners come from\n\nThe chatbot module told you the assistant is made in a second training, where people rate the model's answers and the model is nudged toward the ones they preferred. It did not tell you who the people are. This lesson does, because it is the part of the supply chain that a marketing page never mentions and that the price of a query does not show.\n\n## The work\n\nThere are four kinds, and they pay very differently.\n\n**Labelling.** Marking what is in an image, transcribing audio, tagging a sentence's meaning. This is the oldest kind: the image dataset from the history module was labelled this way, by tens of thousands of people paid a few cents a task on a crowd-work site that has run since 2005.\n\n**Ranking.** Reading two answers to the same prompt and saying which is better, thousands of times. This is the raw material of the second training. Every preference the model has about tone, length and helpfulness was, at some point, a person's click.\n\n**Filtering harm.** Reading the worst text the internet contains (abuse, violence, sexual content involving children) and labelling it so that a classifier can be trained to keep it out of the model and its answers. Somebody has to read it. That somebody is not in the marketing.\n\n**Writing the ideal answer.** Demonstrations of what a good response looks like: a doctor writing the model answer to a medical question, a programmer writing the correct code, a lawyer the correct analysis. This is the expensive end, and since 2023 it has become an industry of its own.\n\n## Who and where\n\nIn January 2023 a magazine investigation found that a leading lab had contracted a data firm in Kenya to label toxic content for its models, with workers paid between about one dollar thirty and two dollars an hour, reading descriptions of violence and abuse for hours at a stretch. Several described lasting psychological harm. The contract ended early. The workers later petitioned lawmakers and formed a union.\n\nThe volume work is done wherever English-speaking labour is cheap and connected: Kenya, the Philippines, Venezuela, and India, through platforms that post tasks and pay per task, sometimes withholding pay for work rejected by a reviewer the worker never sees. One large platform withdrew from Kenya altogether in 2024 with little notice, leaving thousands of workers with no income overnight.\n\nThe expensive end is in the same countries and the rich ones. A platform recruiting people with doctorates to write and grade model answers pays tens of dollars an hour. Volume labour and expert labour are the two ends of the same pipeline, and the gap between them is wider than in most industries.\n\n## India's place in it\n\nIndia is one of the largest sources of this labour, in two shapes. Firms in Kolkata, Hyderabad and elsewhere employ thousands on labelling and annotation, often for foreign labs, at salaries that are ordinary for the local market and low for the value produced. And a small non-profit has taken the other approach: recruiting rural workers to record and label Indian-language speech and text, paying around five dollars an hour (several times the local minimum) and returning part of the revenue from the data to the people who produced it. The datasets it builds are a large part of why models are less bad in Indian languages than they were. It remains an exception.\n\n## Why it matters for what you learned\n\nTwo threads from earlier in the course pull together here. The lesson on where the text came from told you the pile is the filtered residue of the web. This lesson is the other half: the *preferences* that shape the answers are the filtered residue of a workforce, and its composition (its languages, its cultures, its guidelines from the lab) is in every reply. A model that seems oddly American in its assumptions about politeness has learned them from raters working to an American lab's rubric, wherever the raters sat.\n\nAnd the lesson on what a query costs gave you the electricity. Add the labour. A frontier model's second training uses hundreds of thousands to millions of human judgements, and the harm-filtering alone has, at every major lab, been done by people who were not well paid for it. It is not in the token price. It is in the model.\n\n## What is changing\n\nModels now rate other models' answers for a large share of the volume work, which reduces the labour and adds the biases of the rating model. Labs have raised wages and standards at some contractors after reporting, and not at others. Unions and petitions exist and are new. The expert end is growing and the volume end is being automated, which is the module's whole pattern again, applied to the people who built the tool."},"check":{"question":"Why does a chatbot trained by a lab in one country tend to carry that country's assumptions about politeness and helpfulness, even when its raters sat elsewhere?","options":[{"text":"Because the base model's training text is mostly from that country, and the second training cannot change what the pile taught.","correct":false,"tempting":"Names a real influence on facts and register, but the manners come from the second training, not the pile."},{"text":"Because the raters worked to the lab's rubric, so the preferences they recorded carried the lab's norms.","correct":true},{"text":"Because the lab hard-codes politeness rules into the model.","correct":false,"tempting":"Imagines written rules; the model has none, and the manners were learned from ranked examples."},{"text":"Because the raters were all from that country.","correct":false,"tempting":"Contradicts the lesson; the volume raters were largely in Kenya, the Philippines, Venezuela and India."}]},"prev":{"slug":"the-freelance-canary","title":"The first place it showed up in pay","url":"https://addaly.com/learn/ai-actually-explained/the-freelance-canary"},"next":{"slug":"six-trades-honestly","title":"Six trades, as they stood in 2025","url":"https://addaly.com/learn/ai-actually-explained/six-trades-honestly"},"authorship":"Founder and series editor: Dr Mustafa Mun. Written by AI models to a written standard. Exam keys checked blind by a second model: 3,665 of 3,666 agreed. Academic review invited.","review":{"status":"Draft, open for academic review"},"cite":"Addaly (2026) 'The labour that trained it', in AI, Actually Explained. Edited by M. Mun. Addaly. Available at: https://addaly.com/learn/ai-actually-explained/the-hands-that-trained-it (Accessed: 13 October 2026).","licence":{"name":"CC BY-NC-SA 4.0","url":"https://creativecommons.org/licenses/by-nc-sa/4.0/"},"reader":{"schema":1,"title":"The labour that trained it","courseTitle":"AI, Actually Explained","module":{"index":8,"slug":"ai-and-jobs","title":"AI and work, with the evidence"},"answer":null,"keyIdea":"The second training that makes a model helpful and safe is done by people ranking answers and labelling harm, much of it paid by the task in Kenya, the Philippines, Venezuela and India, and the price of a query hides that labour the way it hides the electricity.","hero":{"kind":"plate","image":null,"alt":null,"caption":null},"blocks":[{"t":"h2","id":"where-the-manners-come-from","runs":[{"text":"Where the manners come from"}]},{"t":"p","runs":[{"text":"The chatbot module told you the assistant is made in a second training, where people rate the model’s answers and the model is nudged toward the ones they preferred. It did not tell you who the people are. This lesson does, because it is the part of the supply chain that a marketing page never mentions and that the price of a query does not show."}]},{"t":"h2","id":"the-work","runs":[{"text":"The work"}]},{"t":"p","runs":[{"text":"There are four kinds, and they pay very differently."}]},{"t":"p","runs":[{"text":"Labelling.","b":true},{"text":" Marking what is in an image, transcribing audio, tagging a sentence’s meaning. This is the oldest kind: the image dataset from the history module was labelled this way, by tens of thousands of people paid a few cents a task on a crowd-work site that has run since 2005."}]},{"t":"p","runs":[{"text":"Ranking.","b":true},{"text":" Reading two answers to the same prompt and saying which is better, thousands of times. This is the raw material of the second training. Every preference the model has about tone, length and helpfulness was, at some point, a person’s click."}]},{"t":"p","runs":[{"text":"Filtering harm.","b":true},{"text":" Reading the worst text the internet contains (abuse, violence, sexual content involving children) and labelling it so that a classifier can be trained to keep it out of the model and its answers. Somebody has to read it. That somebody is not in the marketing."}]},{"t":"p","runs":[{"text":"Writing the ideal answer.","b":true},{"text":" Demonstrations of what a good response looks like: a doctor writing the model answer to a medical question, a programmer writing the correct code, a lawyer the correct analysis. This is the expensive end, and since 2023 it has become an industry of its own."}]},{"t":"h2","id":"who-and-where","runs":[{"text":"Who and where"}]},{"t":"p","runs":[{"text":"In January 2023 a magazine investigation found that a leading lab had contracted a data firm in Kenya to label toxic content for its models, with workers paid between about one dollar thirty and two dollars an hour, reading descriptions of violence and abuse for hours at a stretch. Several described lasting psychological harm. The contract ended early. The workers later petitioned lawmakers and formed a union."}]},{"t":"p","runs":[{"text":"The volume work is done wherever English-speaking labour is cheap and connected: Kenya, the Philippines, Venezuela, and India, through platforms that post tasks and pay per task, sometimes withholding pay for work rejected by a reviewer the worker never sees. One large platform withdrew from Kenya altogether in 2024 with little notice, leaving thousands of workers with no income overnight."}]},{"t":"p","runs":[{"text":"The expensive end is in the same countries and the rich ones. A platform recruiting people with doctorates to write and grade model answers pays tens of dollars an hour. Volume labour and expert labour are the two ends of the same pipeline, and the gap between them is wider than in most industries."}]},{"t":"h2","id":"india-s-place-in-it","runs":[{"text":"India’s place in it"}]},{"t":"p","runs":[{"text":"India is one of the largest sources of this labour, in two shapes. Firms in Kolkata, Hyderabad and elsewhere employ thousands on labelling and annotation, often for foreign labs, at salaries that are ordinary for the local market and low for the value produced. And a small non-profit has taken the other approach: recruiting rural workers to record and label Indian-language speech and text, paying around five dollars an hour (several times the local minimum) and returning part of the revenue from the data to the people who produced it. The datasets it builds are a large part of why models are less bad in Indian languages than they were. It remains an exception."}]},{"t":"h2","id":"why-it-matters-for-what-you-learned","runs":[{"text":"Why it matters for what you learned"}]},{"t":"p","runs":[{"text":"Two threads from earlier in the course pull together here. The lesson on where the text came from told you the pile is the filtered residue of the web. This lesson is the other half: the "},{"text":"preferences","i":true},{"text":" that shape the answers are the filtered residue of a workforce, and its composition (its languages, its cultures, its guidelines from the lab) is in every reply. A model that seems oddly American in its assumptions about politeness has learned them from raters working to an American lab’s rubric, wherever the raters sat."}]},{"t":"p","runs":[{"text":"And the lesson on what a query costs gave you the electricity. Add the labour. A frontier model’s second training uses hundreds of thousands to millions of human judgements, and the harm-filtering alone has, at every major lab, been done by people who were not well paid for it. It is not in the token price. It is in the model."}]},{"t":"h2","id":"what-is-changing","runs":[{"text":"What is changing"}]},{"t":"p","runs":[{"text":"Models now rate other models’ answers for a large share of the volume work, which reduces the labour and adds the biases of the rating model. Labs have raised wages and standards at some contractors after reporting, and not at others. Unions and petitions exist and are new. The expert end is growing and the volume end is being automated, which is the module’s whole pattern again, applied to the people who built the tool."}]}],"check":{"question":[{"text":"Why does a chatbot trained by a lab in one country tend to carry that country’s assumptions about politeness and helpfulness, even when its raters sat elsewhere?"}],"options":[{"text":[{"text":"Because the base model’s training text is mostly from that country, and the second training cannot change what the pile taught."}],"correct":false,"tempting":[{"text":"Names a real influence on facts and register, but the manners come from the second training, not the pile."}]},{"text":[{"text":"Because the raters worked to the lab’s rubric, so the preferences they recorded carried the lab’s norms."}],"correct":true,"tempting":null},{"text":[{"text":"Because the lab hard-codes politeness rules into the model."}],"correct":false,"tempting":[{"text":"Imagines written rules; the model has none, and the manners were learned from ranked examples."}]},{"text":[{"text":"Because the raters were all from that country."}],"correct":false,"tempting":[{"text":"Contradicts the lesson; the volume raters were largely in Kenya, the Philippines, Venezuela and India."}]}],"order":[2,0,1,3],"why":null},"prev":{"slug":"the-freelance-canary","title":"The first place it showed up in pay","minutes":8},"next":{"slug":"six-trades-honestly","title":"Six trades, as they stood in 2025","minutes":10}}}