Click Worker

Click workers are people who accept small work assignments through an internet platform and get paid for them. A single task often takes only a few seconds to a few minutes. Typical examples include marking objects in photos, listening to short audio recordings, or evaluating the answers of a computer program. Those who do this work usually have no fixed employment contract, but are instead paid per task completed. The name comes from the fact that the activity consists almost entirely of clicking and short inputs. The clients are often based in wealthy countries, while the workers are spread across the entire world.

The invisible workforce behind AI

A language model or an image recognition system learns from examples. But these examples first have to be sorted and labeled. A computer cannot know on its own which image shows a traffic light and which shows a lamppost. So a human looks at it and enters the correct answer. Millions of such entries together make up the training material used to train an AI.

This means that the success of expensive technology depends on very cheap manual labor. Estimates suggest several million people worldwide work on such platforms, at least occasionally. Well-known locations include Kenya, India, and the Philippines. Reports of wages of one to two dollars per hour have drawn criticism in recent years. Particularly burdensome is the review of violent or abusive content meant to be filtered out of systems.

A common misconception is that modern AI works without humans. The opposite is true: the better a system is supposed to become, the more carefully checked examples it needs. The human element does not disappear — it is simply pushed further back and out of view.

From large order to ten-second task

It starts with a company holding a large amount of data. It breaks the work down into tiny, uniform subtasks. These subtasks end up on a brokering platform, for example Amazon Mechanical Turk or Appen. Anyone registered can pick tasks there and complete them in the browser. Payment is usually in the range of a few cents per item.

You can picture it like a giant assembly line where thousands of people stand at once. Each one sees only their own single step, never the finished product. This makes it cheap, but also error-prone. That is why clients often have the same task processed by several people and take the most common answer. Anyone who deviates too often from the others gets fewer or no more tasks.

It is important to distinguish this from so-called Reinforcement Learning from Human Feedback, or RLHF for short: here, humans rate finished AI responses so that the model subsequently becomes more polite and helpful. These ratings, too, often come from click workers. The difference lies only in what is being rated, not in who does the work.

Where click work shows up in everyday life

The best-known example is the small image puzzles you see when logging into websites. When you click all the fields containing crosswalks there, you are not just proving that you are not a program. At the same time, you are providing a free label for training data. This exact type of task is also assigned for pay on platforms.

In business news, click workers usually come up in connection with two topics. First, in reports about the working conditions of suppliers to major AI companies. Second, in questions about how clean the data is that was used to train a new model. Click work also exists in Germany, for example in checking product descriptions or transcribing receipts. For investors, this is a cost factor: those who have their data prepared more carefully pay more, but get more reliable systems in return.