Imagine a very simple scenario. You open an app, upload a photo, and ask an artificial intelligence system to recognize what's inside. Within seconds, the response arrives: dog, road, traffic light, face, document, prohibited content. It seems like magic. The machine sees, understands, and classifies. We ask a question, and it responds. Everything seems automatic, clean, and instantaneous.
But behind that response are often thousands of people who have looked at similar images before us. People who have labeled photographs, corrected errors, transcribed audio, classified text, and flagged violent content. They are the ghosts of the machine.
The Illusion of Automation and “Ghost Work”
The expression "ghost work" refers precisely to this hidden work that makes artificial intelligence possible. We're not talking about engineers in the labs of large companies, but about a global and fragmented workforce, often precarious, operating behind the interface. When an algorithm seems to learn on its own, someone has actually prepared the world for it.
Recognizing a cat requires a huge amount of images already classified as "cat" or "non-cat." The same goes for moderating toxic content or deciding which response is most useful, least offensive, most coherent. AI isn't born intelligent. It's trained. And training requires human labor.
The removal of labor and the fetishism of the commodity
This point is sociologicamente This is crucial because it shatters the illusion that digital is immaterial. As I tried to demonstrate in the first article, the "cloud" we use every day is made up of servers, energy, minerals, and very real infrastructure.
The problem is that a huge portion of this work is rendered invisible. The user sees only the end result, the company shows off the elegant interface, marketing talks about automation, intelligence, efficiency—but the work that makes it all possible disappears. It's as if the machine presented itself, already prepared, already competent.
Here we return to an old sociological theme: commodity fetishism applied to the machine. We see the finished object and forget the hands that produced it. We see the service, not the workers.
Content Moderation and the Externalization of Trauma
Take the case of content moderators. Every day, platforms must decide what can stay online and what should be removed: violence, pornography, racism, self-harm, misinformation. Sometimes automated systems intervene, but often the people behind it spend hours looking at extremely disturbing images and videos.
I remember the first time I read the testimonies of Kenyan moderators employed by contractors for large platforms: they described grueling shifts spent classifying atrocities, with severe mental health consequences, recurring nightmares, and difficulty returning to normal daily life. The material stayed with me for days. We scroll through a relatively orderly and "safe" feed because someone has absorbed the trauma for us. The promise of automation hides a human delegation: someone, elsewhere, continues to do the dirty work for the machine.
Crowdworking and precariousness of the workforce
Content moderation and microwork aren't exactly the same thing. The former is often organized in outsourced offices, call centers, and delocalized facilities (BPOs). Data labeling and pure microwork, on the other hand, are performed through global crowdworking platforms that fragment tasks and distribute them to isolated individuals, often in their own homes.
In both cases, human labor becomes the hidden infrastructure of automation. An image to classify, a sentence to correct, an audio file to transcribe, a piece of content to flag. Each individual task is worth very little. But millions of tasks, added together, build the behavior of systems we then call "intelligent."
The machine appears autonomous because human labor has been broken down to the point of near-invisibility. This invisibility has concrete political consequences. If we don't see workers, we don't see their low wages, their exhausting hours, their exposure to harmful content, their lack of protections, and the difficulty of organizing.
Ghost work is often sold as flexible: you work when you want, from where you want, and you supplement your income. But this flexibility comes at a high price. The worker is always available, but rarely recognized. It's a new form of structural inequality.
The most advanced artificial intelligence also relies on impoverished and invisible workers, often located in economically vulnerable areas. This isn't an accident. It's a model. The platform needs that labor, but it wants it to be modular, scalable, and available on demand—like electricity: it turns on when needed and off when not.
Unmasking Technology: The Machine as a Social Fact
The term "ghost" is effective precisely for this reason: it's there, it's active, but it mustn't be seen. It's present in the machine, but the audience must believe that the machine operates on its own. Seeing ghosts means demystifying technology: understanding that behind a chatbot are people who have evaluated toxic responses, that a computer vision system relies on annotators who have drawn boundaries around objects and faces, that a fluid translation hides human proofreaders.
Every time a machine "learns," someone has organized the social conditions for its learning. And this work doesn't end when the system is launched: it continues afterwards, with checks, updates, corrections, and adjustments needed when context, language, or social norms change. What we call automation is often, in reality, human labor relocated to less visible points in the chain.
This changes the way we should talk about innovation. It's not enough to ask how powerful a model is or how accurate a system is. We need to ask who paid the social cost of that accuracy, who viewed the worst content, who classified millions of examples while remaining outside the prestige and within the effort.
Sociology and AI
This is precisely what sociology is for: looking where the interface tells us not to look. Thus, the image of AI is changing. No longer an artificial mind suspended in the cloud, but a vast distributed factory, without visible gates but with tasks, timeframes, hierarchies, wages, and inequalities that span continents, homes, platforms, and subcontracts. The intelligent machine is also a social machine. It can help doctors, students, researchers, and businesses. It can make powerful tools accessible. But it can also conceal exploitation and present as automation what is still, largely, human labor.
In the next article, we will delve into this very illusion: the idea that the machine is autonomous, almost magical, when in reality it often functions like a new version of the old Mechanical Turk, the eighteenth-century automaton that seemed to play chess by itself while a hidden human moved the pieces from within.
Bibliography
- Casilli, A. A., & Posada, J. (2019). The platformization of digital labor: How the gig economy shapes the future of work. In M. Graham & W. H. Dutton (Eds.), Society and the internet: How networks of information and communication are changing our lives (2nd ed., pp. 257–274). Oxford University Press.
- Crawford, K., & Joler, V. (2018). Anatomy of an AI system: The Amazon Echo as an anatomical map of human labor, data and planetary resources. AI Now Institute & SHARE Lab. https://anatomyofai.com
- Gray, M.L., & Suri, S. (2019). Ghost work: How images of AI hide a new global underclass. Eamon Dolan Books.
- Irani, L. (2015). Difference and dependence among digital workers: The case of Amazon Mechanical Turk. South Atlantic Quarterly, 114 (1), 225–234. https://doi.org/10.1215/00382876-2831665
