People who wrote the pages→People who labelled them→The company that built it→You, asking
Picture asking a chatbot a question and getting a good answer in two seconds. It feels as if the machine did it alone. Behind the two seconds are four sets of hands: the people who wrote the pages it learned from, the people paid to label and clean them, the company that built it, and you.
Millions of articles
↓→
Learned from them
↓→
Court case, 2023
A chatbot learns from text, lesson 10: billions of pages, most of them written by people for other reasons. Articles, books, forum posts, recipes. Few of those people were asked. In December 2023 the New York Times sued a chatbot's makers for training on its articles without permission. Whether that is allowed is still being decided in courts.
The newspaperwrote the articles
↓→
The companytrained on them
The newspaper's articles were used to train the model. Who decided that?
The readers who asked questions.
Not yet. Readers arrived after the model was built. The pile was chosen before any of them typed a word.
The company that built the model.
Right. The company chose what went into the pile. The newspaper found out afterwards, which is why there is a court case and not a contract.
The newspaper.
Not yet. The newspaper is the one suing. Had it decided, there would be a deal and a payment, not a lawsuit.
Labellers in Nairobi1.30 to 2 dollars an hour
↓→
A model that refuses the worst
Nobody sees the second set of hands. A model trained on the open internet learns its worst parts too. To teach it what to refuse, people have to read the worst text and tag it. From late 2021, workers in Kenya did that for one chatbot's maker, through a contractor, for about 1.30 to 2 dollars an hour.
Slide to the reported take-home pay of those labellers, in US dollars an hour.
Slide the marker to where you think the answer is, then press Lock. Nothing is scored.
12
your guess · 0 dollars an hour
the answer · 2 dollars an hour
01020
dollars an hour
1Five5 dollars an hour
2Fifteen15 dollars an hour
Your guess: —
Between about 1.30 and 2 dollars, by the 2023 report, for reading the worst text the internet holds, all day; some described lasting harm from it. Guesses tend to land higher, because the model's makers are among the richest companies there are. The chatbot's maker paid a contractor about 12.50 dollars an hour for the work. The contractor paid the workers 1.30 to 2.
Computers and power9 pounds of 20 · 45%
Staff6 pounds of 20 · 30%
Profit4 pounds of 20 · 20%
Labellers1 pound of 20 · 5%
Made up and round. The writers' share: nothing.
Where the money goes is also a chain. A chatbot subscription, made up and round: 20 pounds a month. Picture how it splits: computers and power, the company's staff, profit, and the labellers. The people whose writing the model learned from get nothing in most cases today. The labellers' share is the thin slice.
Move the control through the labellers' pay. The labelling needs two million hours; the computer bill is a made-up 100 million dollars.
Labelling2 million dollars
Computers100 million dollars
Pay per labeller, dollars an hour: 1
At 1 dollars an hour, labelling the model costs about 2 million dollars, against a made-up 100 million for the computers that train it. Labelling is a sliver of the bill.
Labelling4 million dollars
Computers100 million dollars
Pay per labeller, dollars an hour: 2
At 2 dollars an hour, labelling the model costs about 4 million dollars, against a made-up 100 million for the computers that train it. Labelling is a sliver of the bill.
Labelling10 million dollars
Computers100 million dollars
Pay per labeller, dollars an hour: 5
At 5 dollars an hour, labelling the model costs about 10 million dollars, against a made-up 100 million for the computers that train it. Labelling is a sliver of the bill.
Labelling20 million dollars
Computers100 million dollars
Pay per labeller, dollars an hour: 10
At 10 dollars an hour, labelling the model costs about 20 million dollars, against a made-up 100 million for the computers that train it. Even here it is under a third of the computer bill.
Labelling30 million dollars
Computers100 million dollars
Pay per labeller, dollars an hour: 15
At 15 dollars an hour, labelling the model costs about 30 million dollars, against a made-up 100 million for the computers that train it. Even here it is under a third of the computer bill.
Even at 15 dollars an hour, labelling costs less than a third of the computers. What does the explore show about the low pay?
Low pay was the only way the model could be built.
Not yet. At 15 dollars the labelling is 30 million against 100 for computers. The model gets built in every frame.
Paying more would change nothing for the labellers.
Not yet. Pay is the one thing that changes between frames, and it is the labellers' pay. What stays nearly fixed is the company's total bill.
The labelling bill stays small next to the computers at every pay level.
Right. Two million hours is a lot of reading and a small line in the budget. Where the pay was set is not explained by what the company could afford.
Move the control.
Each is a set of hands behind a chatbot's answer. Were they asked before their work went in?
One card at a time. Tap the pile it belongs to.
Card 1 of 6
The engineers who built the model.
The forum users whose posts were in the pile.
The labellers who tagged the worst text.
The novelists whose books were in the pile.
The newspaper whose articles were in the pile.
You, typing a question.
The engineers who built the model. → Asked
Not yet. Asked. They were hired to do exactly this, and agreed to it in a contract.
The forum users whose posts were in the pile. → Not asked
Not yet. Not asked. They wrote for other forum users years earlier, and nobody came back to ask about the pile.
The labellers who tagged the worst text. → Asked
Not yet. Asked. They took the job, at the pay offered. Being asked and being paid well are two different questions.
The novelists whose books were in the pile. → Not asked
Not yet. Not asked. Several are suing over it. A published book is not a consent form.
The newspaper whose articles were in the pile. → Not asked
Not yet. Not asked. The December 2023 lawsuit exists because the newspaper found out afterwards.
You, typing a question. → Asked
Not yet. Asked. You chose to type, and you can stop. Your question may also go into the pile, which the small print tells you.
3Asked
vsvs
3Not asked
Right. Half the hands agreed to be there. Half found out later, if at all. The answer on the screen is made from all six.
Writers, not asked→Labellers, paid little→The company→The answer on your screen
So the AI did it hides a chain of people. Some chose and were paid well. Some chose and were paid little. Some never knew. The answer on your screen is made from all of their work. What courts and lawmakers are arguing over now is which of them is owed something for it.
The artistnever agreed
↓→
The generatora picture in her style
An image generator makes a picture in the style of a living artist whose work was in its training pile. The artist never agreed. Which of these is true today?
It is clearly fine, because the picture is new.
Not yet. The picture is new. Whether the copying that taught the model to make it was allowed is the question the courts are still answering.
No court has settled it yet, and more cases are still running.
Right. A British court in 2025 said the model itself was not a copy of the pictures. It did not decide whether the training was allowed, because the training happened outside Britain, and there is an appeal. American cases brought by artists are still going. Nobody can yet say the rule.
It is clearly against the law, and the company will be fined.
Not yet. Some artists believe so and are suing. No court has yet made that the rule. The cases are still being argued.
Lesson complete
A chatbot's answer is made from many people's work. Some chose and were paid. Some were never asked.