Click workers are the trainers of artificial intelligence. They sit in Kenya or Colombia and teach machines about the world. But the work is precarious. Employees report exploitation, lack of prospects and digital surveillance.
There has been a lot of discussion about whether artificial intelligence really intelligent is. It is worth asking whether AI is actually like this artificially – because the technology is essentially based on repetitive work by people, on manual work.
Hardly anyone knows that better than Fred (all names changed). The 27-year-old is a click worker. He lives in Kasarani, a lively district of Nairobi in Kenya. There he sits in a closet in front of his old, panting laptop. Bird’s-eye view images appear on its screen. hundreds of them. Fred uses the mouse to draw lines around a lawn, around a circular swimming pool, and then around a gray tin roof. Then he enters what can be seen: “Roof”, and so on.
Information for AI drones
The images are drone footage from a US logistics company. It delivers packages by drone for retail giants like Walmart. To do this, the AI-controlled drone has to learn where to put the packages: not in a swimming pool or on a roof, but on a lawn.
“We train the drone. Teach it: This is a person, this is the ground, this is water,” explains Fred. Using software, Fred marks and describes the content of images – annotates them, as the technical term goes.
Training data for artificial intelligence
Fred is part of a global industry that processes so-called training data for the development of artificial intelligence. SWR2 knowledge spoke to a dozen of these click workers over several months. In the podcast most of them criticize their working conditions very clearly – sometimes they speak of exploitation. It is her work that makes AI really intelligent.
With the data they annotate – especially images – AI systems learn to recognize signs or traffic lights during autonomous driving, for example. Or intelligent drones to drop packages.
What people can do better
“AI systems are dumber than the hype suggests,” says Milagros Miceli, computer scientist and sociologist at the Weizenbaum Institute in Berlin. To teach an AI, for example, what a cat is, you have to show it as many examples of a cat as possible.” But to describe what a cat actually is in the pictures: that’s what the click workers do.
“The problem is simply that there aren’t enough images that are described precisely enough,” says Kristian Kersting, Professor of Artificial Intelligence and Machine Learning at the TU Darmstadt. Bringing these two “modalities” together, i.e. text and image: A human being can do that much better than machines.
Barely more than a euro hourly wage
People are also needed for AI training using texts. Click workers in Kenya filtered out unwanted responses, such as violent content, for the ChatGPT bot. The goal: The intelligent chatbot should learn from its mistakes and transgressions. The click workers at the AI service provider Sama, who made this possible, were paid almost two dollars an hour, according to research by the Time Magazine made public earlier this year.
Other click workers, on the other hand, are paid even less, like Fred. Like more than 7000 other click workers, he works for the outsourcing company Cloudfactory. Its customers include companies worth millions and billions. However, workers like Fred are paid the equivalent of 1.20 euros an hour. That’s hardly enough at the end of the month, says Fred.
“It’s a bit like slave labor because the people here have no other choice. There are hardly any jobs,” says John, a colleague of Fred’s. Both have studied, but because of the difficult economic situation in Kenya, they only have the badly paid click work for Cloudfactory.
They also criticize the company for another reason: the invasion of their privacy. Because Cloudfactory controls the employees who work from home with digital tools. Using a special browser, the company can not only create screenshots of the workers’ private PCs, but also access their webcams. “This is excessive surveillance,” says John.
German companies also benefit
The German pharmaceutical company Sartorius is one of Cloudfactory’s customers. It has been working with Cloudfactory since 2019. Click workers annotated 5,000 microscopic images of cells for Sartorius, by hand. Datasets like this are used to train AI models, for example to evaluate how cells react to certain drugs.
It is difficult to name such examples of companies that benefited, because the customers of click work, from Google to OpenAI to Tesla, have an interest in the magic of artificial intelligence being preserved, as experts emphasize. “The main advantages of outsourcing this work are obfuscation and scalability,” says AI and digital economics researcher Fabian Ferrari from Utrecht University.
Some bizarre orders
An estimated ten million people work on the processing of training data worldwide. This is also made possible by so-called crowdwork platforms, where almost anyone with an Internet connection can register.
“Digital work is precarious because it can happen 24/7, because it’s poorly paid and faces global competition to the bottom,” says Ferrari. Because the click work takes place mainly in low-wage countries: in addition to Kenya, for example, in Pakistan or in Venezuela.
In conversation with the SWR tell click workers about sometimes stressful orders. John once had to annotate images from an operating room, he says. Another time footage from a porn film. He “labelled” the nude performers: “Male, female, sexual contact one way or the other. Sometimes it was traumatizing.”
Sometimes, he says, he recognizes parts of his work: in a documentary about self-driving cars on YouTube. Another click worker tells how she annotated shots of cut fruit. She later learned by chance that the customer is a company in France that used it to train an AI model to pack fruit salad.
“We need these people”
There are now hardly any industries that do not rely on AI and thus also on click work. The latter is itself changing due to the rapid progress of technological development. But the human work for AI training is not likely to disappear anytime soon, say experts. “It’s a fantasy to believe that it would work without these people,” says Milagros Miceli. “We have to make them visible.”
The click workers in Kenya like Fred want that too. “Companies should value us as much as their AI systems,” he says. Better pay and an end to digital surveillance would be the start.