I had the pleasure of being interviewed by journalist Luciana Leiderfarb for the Portuguese weekly magazine Expresso. Here is the English translation of the original six-page interview in Portuguese.
Interview
Antonio Casilli
AI depends on on millions of poor micro-workers
Investigating how artificial intelligence is produced is at the heart of his work. A sociologist with a background in economics, he wanted to look beyond the end-of-the-world marketing that companies propagate and the “moral panic” that each new technology brings. For him, AI doesn’t destroy jobs: it makes them increasingly informal and precarious. Especially in the Global South
BY LUCIANA LEIDERFARB (TEXT) AND TIAGO MIRANDA (PHOTOS)
In the bustling lobby of a frenetic hotel in Lisbon, where autumnal tourists come and go, laden with suitcases and taking selfies, he is like a fish out of water. Dressed in discreet black, with a pointed beard and John Lennon-style glasses, Antonio Casilli is a sociologist and professor at the Polytechnic Institute of Paris. A 51-year-old Italian, who initially studied economics — what he calls his “first life” — and later sociology, he ended up combining both disciplines, dedicating himself, among other things, to investigating and defining the concept of ‘digital work’ and the impact of AI not on its use, but on how it is transforming and shaping informal work in the so-called Global South. As a co-founder of DipLab (Digital Platform Labour), he is familiar with the terrain and has witnessed how these new labor nuclei, some extremely precarious, are organized, covering, according to World Bank data, 435 million people.
Can we say that your work revolves around digital capitalism?
Yes, my research focus is on the capitalism of digital platforms and how it has changed the way goods and services are produced. It’s a huge paradigm shift that occurred 20 years ago. ‘Platform’ essentially means that we are moving away from a fixed base, like the companies of the 1940s, and transitioning to a new situation where companies are also markets. Think about Amazon: it is both a market and a company. What does it do to work and the services it produces? That’s what I try to answer.
Back in 2015, in a book co-written with Dominique Cardon, you talked about the concept of ‘digital work.’ Why did you become interested in this area?
In my previous life, I studied to be an economist and only later became a sociologist. During that phase, I became interested in the still-emerging digital economy, the internet bubble before 2000, and the associated economy — particularly labor and industrial relations. Ten years ago, I felt the need to revisit everything that had been published in the 1990s. Between 2012 and 2014, I dedicated myself to analyzing this emerging theme of ‘digital work,’ which was at the center of academic interest: there had been a major conference in New York and a book called “Digital Labour: The Internet as Playground and Factory.” At the time, studying how people used the early social media platforms, like Facebook and MySpace, I quickly discovered, along with my colleagues, that digital work was not just entertainment but produced a specific type of commodity, which is artificial intelligence; that clicks and data not only create value for companies but also train AI solutions.
You criticize the apocalyptic notion proclaiming the end of work due to automation. And you say there is a “moral panic” about AI based on wrong assumptions. There is dystopia, but it’s a different kind.
It’s clearly a different kind. This is the theme of a book I published in 2019, “En attendant les robots — Enquête sur le travail du clic” [Waiting for the Robots — an investigation into click work], according to which we should not view automation and AI as destructive but disruptive. What’s the difference? If we destroy jobs, we are destroying employment. People simply become unemployed, and that’s where the ‘moral panic’ that has accompanied us since the early 19th century comes in. The first industrial revolution already came with the terror that people would become unemployed. Not only did it not happen, but the number of formally and informally employed people today is incomparably higher than in previous centuries. Of course, we are 8 billion now, but beyond that, there has been a drastic shift in work. This ongoing automation does not mean mass unemployment; it means the persistence of informal, precarious, contingent, and unprepared work. There are countries where a large part of the population has never had a formal job — and not because robots replaced them. On the other side are the employed people with a real contract, social protection, and respected rights, but they see these rights deteriorating due to constant competition with so-called automated processes. So, this is the starting point. The next question is: what is happening with work?
And what is the response?
Based on years of research in nearly 30 countries across various continents, we have found that work has simply become invisible. The French use the word “invisibilisé,” which means that work is not invisible because we don’t see it, but because someone has made an effort to hide it. And how does it become invisible? Well, in some cases, due to its small and modest nature, inconspicuous. It is not ostentatious. We produce value without realizing it: this is digital work, which is also invisible because it has been outsourced to distant countries for us, Westerners. This is how we come to the necessary outsourcing to produce AI.
Therefore, automation not being the end of work is not necessarily good news.
No, it isn’t. And I’m not providing a comforting counter-narrative. Work will always exist. Robots will never do what humans do. But robots are an excuse, a pretext used by corporate actors to reshape and erode the social rights associated with workers’ conditions.
Before delving into that, I direct the question you started the lecture with in Lisbon: what the hell is AI?
There is a disconnect in how AI is presented by the media, investors, and venture capitalists, as something that completely transcends human experience. We are always thinking of the superintelligent being that AI has eventually become. There is a lot of talk about Artificial General Intelligence (AGI), which is the term used whenever there is a new breakthrough in a technology like, for example, ChatGPT. At those times, investors and the media say, “OK, this is AGI” or “this is a clear step toward AGI.” And what does that mean? It means a creative AI solution, conscious of the context and capable of human cognitive processes. Unfortunately, anyone who has used ChatGPT knows that it is far from being conscious. And what we have seen in recent years, since the commercial boom of AI, are limited AI solutions, like Siri, any transcription software, or facial recognition algorithms, which do not aspire to be general intelligence but only to provide a specific service using machine learning. Basically, it’s a mathematical process that involves teaching the machine by providing it with enough data and examples (N) so that it can be trained. The facial recognition algorithm has to look at billions of examples of faces to one day recognize eyes, nose, and mouth. It learns as it goes, based on examples. Technologies rely on a lot of data, and this brings us back to the question of work. Because, despite the etymological origin of data—something given, coming from nature or the heavens—they are effectively produced. They are commodities produced by ourselves as users, and by millions of people worldwide who perform this AI training work.
Therefore, those who create AI are us.
Yes, “we are the robots,” as Kraftwerk said in an old song.
But this ‘we’ isn’t just the hi-tech engineers and specialists. They are the ones you call micro-workers, right?
When we say we need human work to produce a product, we first think of engineers, data scientists, etc. In other words, highly specialized and well-paid professionals. However, if we look at global numbers and how many of these professionals exist in the world, we’re talking about 28 or 30 million people. On the other hand, very recently, the World Bank published a report estimating the people working on platforms to generate this data, reaching 435 million. This represents 12 percent of the global workforce. A huge number. Even allowing for a margin of error, we understand that this is not a residual phenomenon. We are talking about a significant phenomenon that has been increasing.
What factors are behind this increase?
One of them is COVID-19. Because, with the pandemic crisis, many people worldwide stayed at home without a stable opportunity to earn money and turned to these online platforms or remote micro-work. What we noticed is that some platforms that initially had around 400,000 workers became platforms with 2 or 3 million workers. What kind of work do they do? It’s essential to understand that on these platforms, we find some extreme freelancing. It’s a free freelancing, where you work independently. You go to the platform, sign up, create a worker profile, and start working. And what kind of work do you get? Well, tasks that last a few seconds or, at most, hours, involving watching videos, leaving comments, clicking on links, transcribing sentences, viewing images and detecting objects in those images, filming yourself, downloading or testing an application. These are microtasks, sparsely paid and simple to execute. Officially, significant training is not required to perform them because it’s what we do every day—even children. And that’s why they are so poorly paid. When I say poorly paid, I’m talking about a few cents or sometimes less than a cent. In other cases, more complex tasks are requested, and especially if done in Global North countries, they can be paid a few dollars. According to our colleagues at Oxford, the average pay is at most $2 per hour. So, this is not the job of the future.
Can we talk about a new version of slave labor?
Despite having written a book with the Italian title “Schiavi del Clic” [“Slaves of the Click”], I don’t fully agree with the notion. Because slavery is a historically complex concept. I’m not saying it’s over. Unfortunately, slavery still exists. But this is work based on the choice of free citizens. Of course, the choice is not entirely free or entirely a choice because in certain cases, especially in the countries where we conduct our fieldwork, like Venezuela, Madagascar, or Bangladesh, the economic situation is so terrible that 80% of people live below the poverty line. In a case like this, how free are we? We choose what we can. Now, we are clearly talking about a new layer of the global population that can well be described as the new proletariat. That is, poor workers with very limited possibilities of having a career or better income and, above all, of being recognized for their value. And this is an important point because, in addition to the monetary aspect, these are the workers who are creating one of the most valuable commodities in the current world, AI. They train AI.
Before coming to Lisbon, you were in Madagascar, where you conduct fieldwork on the subject. In how many countries is DipLab [an interdisciplinary research group on Digital Platform Labor] present?
Currently, we conduct research in about 27 countries. Our colleagues at Oxford, who started about ten years ago, focused on Southeast Asia, India, the Philippines, and, of course, Africa—where English is the native language, such as Kenya or South Africa. And we noticed there were issues with Spanish, Portuguese, and French-speaking countries. I’ll give you two examples: Venezuela and Madagascar. Venezuela began, in 2017, to be one of the countries with the highest incidence of micro-work for AI. The reason was the country’s severe economic and political crisis, but also a couple of advantages, such as free electricity and computers, as the Chavez era provided every family with small computers good enough to visit rudimentary platforms and perform simple tasks paid in dollars. What do we see there? We see a country where many people, from their homes, perform remote micro-work, sometimes creating small informal communities, like entire families where the father works in the morning, and in the afternoon, it’s the daughter or grandmother.
And what is the situation in Madagascar?
Madagascar used to be a French colony. Like Venezuela, the level of education is typically high. It has extensive experience with outsourcing companies, such as call centers or tele-services. Many of these call centers have transformed into companies for AI training. Sometimes they are informal setups, in a private house or in the apartments themselves. Access to electricity is difficult due to its cost, and there is not good access to the internet. More importantly, not everyone has a computer. [shows a photograph]. This is a picture of a private house where 120 people work in all rooms, from the attic to the garage. I’ll give you an example of how poorly they are paid: recently, we entered a house where a company working on ChatGPT-related tasks operated, boasting about paying workers well—around 120 dollars. However, at the entrance, there was a barometer showing what workers earn for each goal they achieve. In this barometer, earnings were indicated with different quantities of beans, not in money. According to some respondents, this was indeed how they were paid, which is not good. Most of them earn just enough to buy a bag of rice.
The geography of this picture reflects global inequalities and those of colonialism. The North against the South.
Let’s be clear: when it comes to inequalities and dependencies, everyone in the scientific community agrees that we have a problem. Higher-income countries, like Western ones, but also China, India, and now Russia, are increasingly becoming the countries where AI is marketed. But these are not the places where it is produced. It is produced in low-income countries facing significant challenges. On the other hand, we clearly see some colonial persistence: these data markets to train AI are organized based on linguistic criteria. Spanish-speaking countries tend to work with Spain or the United States, French-speaking Africa works with France, and many former English colonies, notably India and Bangladesh, work with the UK or the US. The process is even more complex because emerging countries like Russia attract micro-workers from Venezuela without a colonial replication. When I was in Egypt in 2020, I interviewed people working for Chinese platforms created to train China’s surveillance and facial recognition algorithm. And there is no colonial dependence between Egypt and China.
Can this phenomenon be called a kind of ‘low-cost science,’ where the essential element for developing AI is almost cost-free? What ethical questions does this raise?
Firstly, AI is not just science; it is now a significant market and a massive commercial phenomenon. On the other hand, low-cost science is primarily carried out by universities due to funding constraints. However, Google, Meta, Microsoft, and Amazon do not face these constraints. They only aim to encourage and favor their investors, not their workforce. This is basic Marxism. Instead of distributing revenues between capitalists and workers, companies tend to pay the latter less and less. Because they do not recognize the real value provided by the millions of digital micro-workers on which AI relies and depends, who exist globally. This is an old story, the foundation of global capitalism, which means producing in places where labor costs are lower. From this perspective, we face an ethical problem that goes far beyond what we typically associate with AI. Until now, we have focused on how AI is used. But no one seems interested in the phase that precedes usage, in how AI is produced. We should ask what companies do when they produce AI. Do they exploit workers? Do they destroy environmental resources? Do they disrupt legal or political systems? These are scary and important questions.
There are several specific terms associated with your work. Has a new lexicon had to be created to understand AI?
We use many terms because it is still a relatively young scientific community. Probably, in 20 years, we will all speak of one thing. Twenty years ago, one of the first expressions developed was “click work.” That is, we can have AI solutions through the work of people who are not experts, scientists, or computer scientists, and who perform tasks as small as a click. Then, another term had to be invented to cover the quantity issue because it’s not tasks given to ten people. Hundreds of thousands of people are needed for each AI solution that one wants to develop. So, we started talking about “crowdwork.” But complex tasks can also be given to a crowd. Let’s make sure we agree that these are small, simple tasks that are easy and quick to perform—thus, the term ‘micro-work’ emerged. More recently, the situation has become more complex because even microtasks are performed for a variety of reasons and in a variety of ways. For example, in the case of poorly paid workers in low-income countries, these tasks only serve to generate data and are usually associated with so-called “click farms.” The idea of a farm evokes something dirty and close to the ground—it is the lowest level of this market. Now, there is also talk of “AI labor,” generally associated with human contribution to AI solutions.
Could you please describe in more detail the type of work that is digital micro-work?
A significant revolution that happened in data science is that data ceased to be merely values and began to be many different things. An image is data, a piece of data. A video, a conversation, or a snippet of conversation can be data. Let me give you a simple example: a photograph of a salad. This is an example from our fieldwork. Faced with the image of a salad, we have to teach an algorithm to recognize the different ingredients of the salad. The worker has to draw boxes around all the ingredients, like tomatoes, feta cheese, etc., and add a small label, a hashtag. These data—the image, the boxes, and the labels—are recorded in a database similar to an Excel spreadsheet and used to teach the salad recognition algorithm or a self-aware AI what the different ingredients are.
And this microtask is ‘macro’ in terms of the quantity of workers who do it, right?
Yes, because thousands or millions of examples like this need to be produced for machines to learn what the different types of salads are. Other examples that come from our experience are tasks related to voice. Workers receive three-second conversation snippets, probably recorded by a Siri. Even this conversation we are having here can be monitored—indeed, we know it is. The worker has to compare the automatic transcription provided by AI with what they hear as a native speaker. They receive an excerpt of a conversation, the provisional transcription, and have to verify if the spoken word corresponds to the transcription.
ChatGPT itself contains in its acronym this process, right?
The ‘P’ in ChatGPT stands for “pre-trained.” This means that this AI has been pre-trained to do what it does, generate text, over several years and based on a truly enormous amount of data. Huge amounts of data were collected by another platform called Common Crawl, which collects raw data from the internet. The role of micro-workers was to confirm: “OK, this book is about science in the 17th century.” Or “this book is about the pirates of the Caribbean.” We are talking about thousands of annotations made by thousands of workers. When we checked where these workers were located, surprise: they were in countries like Turkey, Kenya, South Africa, the Philippines. So, we were back to square one. Invoicing requires a lot of unpaid and invisible human work.
Do you use ChatGPT?
I started using it because I teach sociology to engineers and knew that my students would use it. I decided to try it with them to help them use it appropriately and academically responsibly. First and foremost, I wanted to teach them not to trust ChatGPT because it is not a factual tool. It exists to provide a kind of creative text that it just invented, a fabrication. But when we use it to write a report for a class, tragedies like false references or untrue arguments are likely to occur. So, I thought the first thing to do was to teach them where ChatGPT goes wrong.
So, is it alarmist to say that it will replace, for example, journalists?
Maybe. Some pedagogy needs to be done, just as with any innovation—think of Google or Wikipedia. We know that students use Google or Wikipedia to search for information; the question is not to use them uncritically. There is a lot of false information, and it’s not just about ‘copy and paste.’ The same applies to ChatGPT: whenever a tool like this is launched in the market, part of the marketing campaign is based on pure terror. It’s fear-based marketing, where the tool’s producer has to show that it is so impactful that it can even be dangerous.
Is it a matter of propaganda?
Curiously or ironically, propaganda is dangerous for categories of workers and professionals who can be competitors of the company producing ChatGPT. What can be done with it? It can be used for teaching, producing texts, and some creative work. So, ChatGPT supposedly threatens education as well as media and journalism, areas already quite disrupted. I always give the example of the electronic calculator. Thirty years ago, math teachers considered that it was also the end of the world because students would never remember on their own what the square root of 7 is. But who needs to know that? The calculator allows me to focus on other things. When it comes to jobs that are said to be under threat, I hope it will be shown that these tools are not so revolutionary, that we still need human creativity, accuracy, and judgment. And that will only happen if we look at ChatGPT for what it is: no more than a text management tool, an auto-fill tool.
Finally, the ‘million-dollar question’: how do you regulate this enormous and profitable market?
Efforts are being made worldwide for AI regulation. The EU has been working hard in recent years to reach a package of 30 new directives on data, AI, algorithmic management, etc. In the US, we have Biden’s new executive order on AI, which is a way of saying that the country is concerned about it. India and China have been creating their own regulatory framework. The problem is that it is always an effort made afterward, which means that technology that already exists tends to be regulated without intervening in the production process. We have to ask ourselves how AI is produced and with what ingredients. We have to discuss the material part of AI, for example, whether it is disposable and whether lithium dependence does not destroy too many resources. We have to talk about data and micro-work, and create rules that can be applied in both the Global North and South—which is complicated but not impossible. Basically, knowing whether a German company recruiting micro-workers in the Philippines will pay the Filipino or German minimum wage. Or if, faced with a dispute between the platform and the company, which court will decide who is right or wrong. Or yet, which country’s law will apply in the event of a strike by these workers. We must ensure that human rights, safety, and protection are respected throughout the entire supply chain, not just in the Global North. As you can see, we can list an endless number of questions. There is much work to be done.