Annotating with care, for care: the real labour behind AI

3 min read

One afternoon, Zhang Na was processing a massive batch of satellite data for an autonomous driving project. On her screen, among the vast mass of digital coordinates, she spotted a familiar shape. It was a distinctive road sign on a road just minutes from her hometown. She immediately took a photo of the screen and sent it to her team's chat group. »This is exactly the road... [near] my home«, she told her colleagues. In that moment, the abstract digital map collided with her physical reality. She realised that if she made even a small mistake here, it wouldn’t be just a glitch in an algorithm; it could mean directing a neighbour down a dangerous path out in the real world. 

Zhang Na is a data annotator in a rural county in Northwestern China. She is part of the large invisible workforce that labels information – text, images, audio or video – so that AI systems can learn to understand the world. For the past decade, this workforce has been the subject of intense critique, as scholars and activists have exposed the »digital sweatshop« conditions in the Global South.

More recently the narrative has shifted, however. With the advent of Generative AI, the industry has promised to »fix« these labour issues through automation. Headlines have been dominated by such publicised moves as Elon Musk’s xAI shaking up its data teams to prioritise efficiency, or companies such as Sama launching Bulk Annotation tools that promise to reduce human effort by 80 per cent. The prevailing logic (or narrative) is clear: we are moving towards a future in which AI builds itself, downgrading or even eliminating the »messy« human element.

However, this automation myth overlooks a critical reality. As Zhang Na’s experience shows, annotation is not just mechanical »click work« that can easily be automated away. It is a complex form of meaning-making that requires navigating constant ambiguity.

To understand the true backbone of AI, we must look beyond the hype of automation and the binary of exploitation. Instead, I argue we should recognise the caring labour behind this work. Annotators such as Zhang Na are not just processing data; they are embedding culture and human experience into code. They annotate with care, using ethical judgement to manage toxic content and ensure safety, and they annotate for care, using this digital labour to sustain their families and communities. In a rush to automate, the industry risks losing the very human infrastructure that makes AI safe and usable.

Annotating with care: the burden of responsibility

Tech companies often portray data annotation as mindless clicks performed by invisible hands until the algorithms become smart enough to take over. However, the data workers I spoke to described a role that requires intensive »response-ability«, with an ethical obligation to the humans who will use these AI systems. This raises a critical question: given that such work is often criticised as tedious and exploitative, shouldn’t we welcome the advent of a kind of AI that might be able to automate it away? 

But things are not that simple. While AI can handle rote tasks, it lacks the cultural intuition and »social responsibility« that workers such as Zhang Na provide. Since starting her job, Zhang Na admits she has developed a heightened sensibility to her surroundings. When walking down the street, she now obsessively checks for traffic regulations and construction zones. 

»You need to have a sense of social responsibility«, she explained. »If you provide incorrect information... it could result in someone being directed onto a road that is impassable. «According to Zhang Na, while AI can identify a road, it often tasks humans with understanding the consequences of, for example, sending a neighbour down a dangerous road.

Such labour demands meticulous attention to detail and patience, traits that traditionally are feminised and undervalued, even though they are significant in relation to safety. Far from being a kind of »human robots« on a digital assembly line, these AI data workers insert human judgement into systems otherwise designed to develop maximum speed. In the rural county where I conducted my research, this workforce is significant. Over 70 per cent of the annotators are women from nearby villages. Many began with only basic computer literacy and have now become certified »AI trainers«. Along with workers with disabilities and vocational students, they have found a precarious but vital foothold in the digital economy, constantly exercising this »response-ability« to bridge the gap between messy reality and rigid algorithms.

Annotating for care: the filter for the world

If annotating with care is about accuracy, annotating for care is about protection. Before AI models can be released, these data workers must be trained to filter out toxicity. Currently, this means that human workers act as a filter, absorbing violence, hate and pornography so that users don’t have to.

The psychological toll of this »infrastructure maintenance« is immense. Li Wenyan, an annotator assigned to label audio pornography, described the task as »oppressive«. In our interview, she stuttered, noting that the repetitive listening made her and her colleagues feel »physically unwell«. Another worker, Wen Tian, was assigned explicit audio tasks while pregnant. The voices made her physically sick, but she was told that few other tasks were available, and eventually she was forced to take low-paid leave. These workers are cleaning up the digital environment so that others may use it more safely, often at the expense of their own well-being.

This is where the debate about AI automation needs to shift. Instead of asking whether AI will replace these jobs, we should ask how AI can protect these workers. AI can – and should – be deployed to absorb the first layers of toxicity that have long fallen on the shoulders of data annotators. By using AI systems to pre-filter the most traumatic content, we can reduce human exposure to psychological harm while still maintaining necessary human oversight for tasks requiring nuance, fairness and contextual judgement. Beyond this protective layer, AI also offers opportunities to address long-standing critiques of the industry. For example, it can help to automate repetitive labelling tasks, improving working conditions. It may also bring to the surface bias patterns at an early stage so that annotators are not blamed for structural issues. And it may enable more transparent workflows in which annotators’ labour, expertise and cultural insight are better valued. In short, AI doesn’t eliminate the need for human annotators; rather, it can be used to redesign the ecosystem around them, shifting from extraction and invisibility towards protection, dignity and recognition.

The »human secret« of AI is not just that humans are still in the loop. It is that the loop is held together by care. Whether such workers are rural women, vocational students or people with disabilities, they are not passive victims. They annotate with care to ensure public safety, annotate for care to filter digital toxicity, and build communities of care to survive. As we look toward the future of work, we must acknowledge this invisible labour and ensure that the AI trainers building our digital world are treated with the dignity and protection they deserve.

The Inclusive AI Lab at Utrecht University is dedicated to incubating leaders and helping to build inclusive, responsible and ethical AI data, tools, services, policies and platforms, with a special focus on the Global South.

Guanqin He is a PhD student and researches AI and work at the Inclusive AI Lab at Utrecht University. Her research focuses on platform gig workers, gender and AI, and migration in China. Guanqin’s work has appeared in journals such as First Monday, Big Data & Society and Mobile Media & Communication.