Recension Society

Cannibal Machine

About: Antonio Casilli, En attendant les robots, Seuil

Are intelligent machines attacking human labor? In his study of the digital labor hiding behind the promises of automation and robots, Antonio Casilli notably argues that social media constitute a form of unpaid work.

Does liking or sharing an article on Facebook or watching a video on YouPorn constitute work? Yes, replies Antonio Casilli in his simultaneously dense, well-documented, and stimulating book En attendant les robots: enquête sur le travail du clic (Waiting for Robots: An Inquiry into Click Work).

This sort of work that does not look like work is what Casilli—along with others—refers to as digital labor, one of the key concepts of this book that might be described as a critical sociology of artificial intelligence (AI). The author, himself a sociologist of work and researcher at the EHESS, suggests that we move beyond AI washing and hype to better see the human, all too human dogsbodies that work in the shadow of AI algorithms.

According to Casilli, the mistake would be to believe that work will disappear with automation. Work will instead be « platformized » and digitized. In other words, while we wait for robots—which, as in Beckett’s Godot, may be slow to come on the scene—we would do well to examine the concrete changes taking place in human labor. For instead of the replacement of work by robots, what we are often dealing with is mere displacement. The example of ATMs is a case in point. These do not completely replace the work formerly performed by bank tellers: rather, it is customers who are now responsible for operating the machines. In many ways, this example is symptomatic of our difficulty in perceiving the new forms of work analyzed by Casilli.

Digital Labor

The concept of digital labor primarily designates the work of the finger, “which is used to count, but also to point, click, press the button” (p. 48). However, things are also more complicated than that. This is the subject of the central section of the book, which continues the debate initiated with sociologist Dominique Cardon on the relevance of this concept. Located in a “grey area between sub-contracting and salaried employment,” digital labor is part of “a relationship of subordination, surveillance, and unequal rights between workers and owners of digital services” (p. 267). It can be classified into three forms.

Labor on-demand is the first form of digital labor. Transport or delivery services such as Uber or Foodora rely on platforms and mobilize decision-making processes based on algorithms and data that disrupt the organization of work. The point is somehow to push the Fordist-Taylorist (assembly-line work) logic to its limits via the increasing measurement and division of labor. And workers do not seem to be the winners here.
Micro-working, which is the subject of the fourth chapter, is the second type of digital labor. It involves the performance of “small chores such as annotating videos, sorting tweets, transcribing scanned documents (...)” (p. 119). This focus placed on the work of these often-precarious people—a hundred million people in the world, including 260,000 in France—who keep digital capitalism running is undoubtedly one of the most interesting contributions of this book.

For instance, online services such as Mechanical Turk allow companies to outsource “click work.” This Amazon platform, which was created in 2005, employs mainly American and Indian citizens who perform poorly paid micro-tasks. These individuals work at home, do not meet each other, and do not know their customers. Obviously, they are not unionized or even organized. There is indeed a lot of irony in the fact that this “artificial AI” service, as Amazon’s CEO Jeff Bezos calls it, is named after the Turk, the 18th century automaton that pretended to play chess, but actually concealed a human being behind its useless gears. How can one not perceive a disturbing echo with the current situation?

As Casilli sums up, “automation comes down to a simple formula: a façade with an engineer boasting about the prowess of his or her machine and a backroom where workers are killing themselves over micro-tasks” (p. 136). Invisible micro-working also involves new forms of control, since customers need to be assured that it is being carried out by humans. Accordingly, the latter must complete CAPTCHA-type tests every forty-five minutes. More generally, micro-workers (turkers) must share their data to operate the platform and be put in contact with customers—with the platform ultimately earning an “intermediation income.”

As you have probably guessed, sharing an article on Facebook or watching a video on YouPorn is neither labor on-demand nor micro-working, since it is neither a paid activity nor a service requested by a customer. But there exists, Casilli continues, a third form of digital labor, networked social work, which notably corresponds to user participation in social media. How can we analyze this phenomenon? The fifth chapter details the controversy over the exact nature of this “free labor.”

Free Labor?

From a “hedonistic” perspective, social media do not fall under the category of work, because people participate in them for free and for pleasure, through a new culture of amateurism. Casilli recommends instead a labor-oriented perspective that views “produsage”—i.e., activities that intertwine production and usage—as part of an “exploitative relationship” (p. 168). Unpaid labor is not new: domestic labor, volunteer labor, forced labor, etc. But how can simple clicks on social media be treated as work? By acknowledging that in creating data (and metadata), the user is creating value. When liking or sharing an article on Facebook, one enriches a database that will allow the company to generate substantial advertising revenues—these amounted to $55 billion in 2018.

Casilli shows that the line between professionalism and amateurism is often a porous one. Is there really a difference in nature between the food critic, the YouTube chef, and the anonymous Yelp reviewer? Does the moderator paid to filter violent content do anything other than the simple user who blocks, mutes, or reports such content? Some may reply that work is defined primarily by its tedious nature. Well, then, why should the fact that a task is pleasant, convivial, or creative disqualify it as work? Besides, are we so certain that social media is free of drudgery? The author cites field studies in which young people from working-class backgrounds describe the discomfort and “drudgery” of being on Facebook “to control what others expose about themselves” (p. 187).

Yet, what seems to be the most decisive argument is the existence of “link farms” and “click farms.” Thus, fake clicks on Facebook are reported to have generated 200 million dollars in 2013 via services like GetPaidForLikes. Facebook itself now offers users the opportunity to “boost a post” for a few dollars. But if people are remunerated in “farms” to perform the same kind of micro-tasks as those executed by ordinary users, should we not conclude—by analogy—that the latter also constitute work?

Thus, when you like or share an article on Facebook, you are helping give it a visibility that might otherwise have been purchased. While the value of this work is tiny (Casilli explained in an interview [1] that one click was worth $0.008 in India in 2019), its multiplication and aggregation creates substantial wealth—provided that a company is able to value the data. “On social platforms,” Casilli concludes, “every new subscriber is now caught up in a click-generating system based on invisible labor—the “free” labor he or she is driven to provide and that of his or her micro-paid counterparts” (p. 218). In other words, despite its playful and free nature, click work is unquestionably part of the digital economy.

Data Valuation

If we find it difficult to view this work as genuine, it may be because it is easier to understand the value attached to an object than that attached to a database or an information flow. To clarify his point, Casilli proposes to distinguish three forms of data valuation—i.e., three ways of adding value to data.

First, one can qualify a content—i.e., add information—for instance by creating a user profile, posting a comment, or liking a video on a social network (it is estimated that Facebook users generate 4.3 billion likes per day). One can also qualify another user—and activate reputation mechanisms—as is the case with dating applications (when one is swiping left or right on Tinder).

One can also create value by monetizing personal data. This is the main business model used by platforms that sell targeted advertising. Databases are composed of cross-referenced and sometimes unexpected information: phone or computer brands, location, contact lists, online behavior. Data brokers are the direct agents of data valuation through monetization.

With the development of AI systems, a third form of data valuation has emerged: valuation through automation. Indeed, deep learning algorithms must be trained with large amounts of “personal” data to make optimal predictions (AI is essentially a prediction tool). Casilli gives the example of controversial facial recognition applications, but one could argue that most of today’s well-performing AI systems—from Google translation to tumor recognition systems—are based on data valuation through automation.

In this reading, then, humans who label—whether or not they are remunerated—appear to be the main source of data valuation. Consequently, their concealment behind the prowess of AI and the promise of robots supports the thesis that the invisibilization of human labor is central to the digital economy. Yet, one can wonder whether automation based on “non-human” data (derived, for instance, from animals, robots, meteorological events) might not undermine these analyses. Likewise, what will happen to this model if, as AI researchers hope, unsupervised learning—i.e., without labeled data—develops?

A Social Critique

For now, on can hardly deny that Casilli has put his finger on an important dimension of the contemporary economy and sociology of digital technology. Other books, such as the recent Ghost Work: How to Stop Silicon Valley from Building a New Global Underclass by Gray and Suri (2019), point in the same direction. Not surprisingly, Casilli’s book ends in the final section with a social critique. After all, “micro-, under-, poorly- or unpaid work appears as a common thread that unites the different forms of digital labor” (p. 223). Thus, En attendant les robots seeks to promote the recognition of digital labor with a view to creating real class consciousness among click workers.

To put a face on these workers, Casilli mentions the short documentary “The Moderators” (2017), which is aimed precisely at showing them to us. In an Indian classroom, a one-week training for future content moderators is being taught. For most of the trainees, this is a first job which they welcome with joy. Despite the sounds of the street, they listen attentively to the instructions of the sari-clad trainer: ”Why don’t you see pornography when you’re on Facebook,” she asks. “Do you think it’s because no one posts pornography?” This is what content-moderation work is all about: To make sure that Facebook does not become YouPorn. And if this documentary nicely complements En attendant les robots, it is because it shows this very prosaic reality: images of penises to be rejected, large ceiling fans, shirts with rolled-up sleeves, and eyes focused on the screen.

While we wait for robots, AI and the digital economy will continue to require Indian moderators, American turkers, and the data we create, often unbeknown to us, on social networks. For all of this produces value. In his conclusion, Casilli explores some avenues of redistribution, with a focus on the idea of digital basic income or on Trebor Scholz’s studies of “platform cooperativism” and collective ownership of the means of digital production.

All these elements make this pleasantly readable book an important contribution to the critical sociology of AI. Casilli no doubt offers a refined and richly illustrated description of digital labor. He also proposes analytical tools and a perspective engaged in the normative debate. Regarding this last point, there is obviously room for discussion. What, for instance, of the well-being associated with click work?

In the case of networked social work, Casilli seems to overlook the value of the free service provided by platforms like YouPorn or Facebook. Whether one is masturbating or informing oneself, there is little doubt that value is produced. And even if this is at the cost of sharing information, it is still a good deal. One might also think that sharing a review of an interesting book will produce epistemic value. And, of course, it is always good for the ego to receive attention, love, or likes.

Antonio Casilli, En attendant les robots, Seuil, 2019, 400 p., 24 €.

by Martin Gibert, 23 January

To quote this article :

Martin Gibert, « Cannibal Machine », Books and Ideas , 23 January 2020. ISSN : 2105-3030. URL :

Nota Bene:

If you want to discuss this essay further, you can send a proposal to the editorial team (redaction at We will get back to you as soon as possible.


[1For the Brasilian journal Epoca, 20 December 2020.

Our partners

© - Any replication forbidden without the explicit consent of the editors. - Mentions légales - webdesign : Abel Poucet