Have you ever generated an image using Large Language Models (LLMs), such as Gemini or ChatGPT? They produce new images and documents in the blink of an eye, when it would have taken several hours and specialized abilities if it were for us to make those. Therefore, many assume that the data learning and collecting process are automatized, unable to see the human labor that lies behind them. But that is not the case. Browsing the Internet, you might have stumbled upon unsettling, harmful images generated by LLMs. To prevent this from happening, there is a process called data labelling, also known as data annotation, which is manually done by human workers. Data labelling helps generative AI detect certain images, determine which answer it provides is more appropriate and useful, and rule out violent, hateful, and sexual data from its database. However, controversy has risen ever since big tech companies started to outsource this task to low-wage workers in developing countries.
LLM models are created through multiple processes. First, massive text data is accumulated from multiple contents including books, web pages, and dialogues. This collection of data then goes through data filtering and cleaning which removes inappropriate or low-quality content. Such data is then tokenized into small units of words, and the models are trained to predict the next token to come after some input. Then comes post training, where data annotation plays a substantial role. Human-in-the-loop data workers help teach the models how we want them to behave by evaluating answers and classifying data.
As the demand for data labeling and human feedback has grown, a new type of global industry has emerged to provide the labor required to train AI models. Since directly hiring individual workers to perform the task comes as a huge burden for AI developers, many developers outsource parts of this work to specialized data annotation companies and online platforms. Since most of the data labeling and classification can be done online, workers do not need to be in the same physical space or country as AI developers. This has allowed business process outsourcing (BPO) firms to make use of global wage differences, hiring workers specifically from countries where labor costs are comparatively low. The way these workers are employed significantly differs as well. Some data workers are directly employed by BPO firms, while others work as independent contractors through online platforms. This provides flexibility of recruitment to outsourcing companies, rapidly expanding the workforce when there is higher demand and reducing when demand falls. For such workers, however, this means unstable and less predictable income and employment status. The major issue of this industry is not the outsourcing structure but the lack of proper protection and adequate compensation for the workers behind the huge AI supply chain.
In 2024 and 2025, several lawsuits were filed against Scale AI concerning the improper working conditions of its data workers. One of the major issues in the lawsuits was the misclassification of workers. Data workers were classified as independent contractors, whereas in reality, they were more like employees since Scale AI controlled most of their working conditions. Because they were treated as independent contractors, they were unable to receive basic work rights such as minimum wage despite working overtime. Another main concern was that although data labelers were constantly exposed to violence and disturbing content in the workplace, there was no psychological protection or support. Yet, Scale AI has denied such allegations. It claims that it legally classified the workers as independent contractors and does not have the duty to apply wage-and-hour laws to them. Regarding the controversy over unpaid wages, it maintains that it complied with labor laws and paid proper wages. Scale AI responded to the exposure to harmful content allegations with claims that it has been operating several safeguards such as notifying workers of disturbing contents before labelling, allowing them to access mental health and wellness programs, and setting an opt-out option if workers no longer want to participate. AI automatically generates new texts and images. Yet, AI itself is not created automatically. Behind the glorious AI industry, human workers are being exploited. As AI models are rapidly developing, there is an urgent need for more transparency, work ethics, and new legislation for those behind.
