Exploited Data Workers Are “the Secret Ingredient of AI Itself”

Despite being branded as autonomous and self-sustaining, artificial intelligence relies heavily on human labor — an army of data workers all over the world, primarily in the Global South. These workers “are purposely hidden by the companies that produce AI because they are the secret ingredient of AI itself,” says sociologist Antonio Casilli, the author of Waiting for Robots: The Hired Hands of Automation and co-writer of the documentary In the Belly of AI. Casilli estimates that over 160 million people are involved in data work around the world.

TRANSCRIPT

This is a rush transcript. Copy may not be in its final form.

NERMEEN SHAIKH: This is Democracy Now!, democracynow.org, The War and Peace Report. I’m Nermeen Shaikh.

One of the most discussed aspects of how artificial intelligence will transform society is how it will impact jobs and labor. On Wednesday, Bill Gates published a nearly 6,000-word essay in which he warned that society is not ready for the transformation and upheaval AI will cause. In the essay, Gates calls AI’s impact on the labor force one of the three primary risks of the technology, saying many jobs will disappear forever and that, quote, “AI will either be the greatest equalizer ever invented, or the worst source of injustice.”

But what gets much less discussion is how artificial intelligence has already transformed work globally. Our next guest is sociologist Antonio Casilli, author of Waiting for Robots: The Hired Hands of Automation. He writes, quote, “Human workers make automation possible. Recognizing the value of their contributions to these digital infrastructures is the first step toward worker reparation and rebalancing the current power dynamics.” He’s also co-writer of the film In the Belly of AI. This is the film’s trailer.

SIMS WITHERSPOON: AI can be a transformational tool in our fight against climate change.

MARK ZUCKERBERG: We’re really optimistic about the potential for AI to help scientists cure, prevent and manage all diseases in this century.

SAM ALTMAN: Some robots will eat some jobs, and some robots will kill humans. But when you look at the trend as a whole, I think it’s going to be incredibly positive for humanity.

UMA RANI: All kinds of artificial intelligence need humans behind them, an army of workers.

MILAGROS MICELI: There is a report from the World Bank in which they actually argue that there are around 150 million data workers and 430 million data workers around the world.

UMA RANI: The work is invisibilized, and the workers are invisibilized.

MILAGROS MICELI: But what are they trying to hide?

FAUSTINE MAKIRA: Can we take a break, please?

INTERVIEWER 1: Yes, of course.

INTERVIEWER 2: Yes, of course.

ANA VALDIVIA: A cloud is not a cloud. It is concrete, right? So, it’s cement. The carbon footprint, carbon emission of a data center is huge, because most of the electricity that they use comes from fossil fuels.

INTERVIEWER 3: There is one AI company which says that it usually removes people from poverty. In your opinion, is that true?

KENYAN DATA WORKER 1: No, no, no.

KENYAN DATA WORKER 2: It’s manipulation.

KENYAN DATA WORKER 3: It’s just a punchline they use to attract people who are — who are qualified.

UMA RANI: What is this AI development? How does it really help our society and the development process? Is it something that we really need?

NERMEEN SHAIKH: That was the trailer for the film In the Belly of AI, directed by Henri Poulain and co-written by our next guest. Antonio Casilli joins us from Paris. He’s professor at the Institut Polytechnique de Paris and author of Waiting for Robots.

Professor Casilli, welcome to Democracy Now! Thank you for joining us. If you could begin by talking about these data workers, workers around the world whom you’ve met who are automating AI?

ANTONIO CASILLI: Yes. Data workers are today’s factory workers insofar as AI has become like a driving force for so many industries, and they are necessary to produce what we use every day — you know, chatbots and models. Now, we are not talking about tech workers, engineers, product managers, designers, these highly paid and highly recognized professionals. We are talking about persons who are hidden. And I insist on the fact that they are not invisible. They are purposely hidden by the companies that produce AI because they are the secret ingredient of AI itself. They train it. And this is a verb that we are now more familiar with, because we know that, I don’t know, ChatGPT has been trained for a number of years by persons who were literally feeding them data and enriching this data. But they are also, these data workers, those who verify that, for instance, the results of the same ChatGPT are in line with what we are looking for as users. And sometimes they are even those who, you know, operate maintenance, perform maintenance live, meaning that sometimes they replace the machine itself, and so they, to an extent, impersonate AI. So, they do all these things, and they are really poorly paid.

We have been following and working with them for a number of years, for six years now already, and we have been conducting surveys and interviews and observations in more than 30 countries and interviewed more than 4,000 of these workers in places like, I don’t know, Venezuela or Kenya or Madagascar and so on. And they are really everywhere. They are also in the U.S. and Europe, also in high-income countries. But, of course, given the very low wages that they have as a compensation, these kind of jobs, which are not jobs — these are more tasks that they perform in a more or less formal way — they are more interesting to people who are living in low-income countries.

NERMEEN SHAIKH: And if you could explain, Professor Casilli — I mean, one of the things that’s so shocking, because indeed these workers are hidden, if not invisible, as you say, just the sheer number of workers there are. I mean, the World Bank report found that there were anywhere between 130 and over 400 million workers, gig workers. I’m not sure how many of those gig workers are data workers. But if you could explain also, I mean, the very curious — because you’ve been to Madagascar, you’ve been to Kenya and various other places, including Finland and Bulgaria — the way in which these workers are compensated? In some cases, as you’ve said, in Madagascar, they’re not paid in money, but in sacks of rice or sugar.

ANTONIO CASILLI: Yeah, sure. Yes, this is something which is appalling, which is, unfortunately, very common. First of all, let’s focus on the estimates that the World Bank has been producing for a number of years, because they started, you know, conducting this kind of service and trying to estimate the number of people who perform online gig work. We are not talking about gig workers like, I don’t know, Uber drivers or delivery persons, you know, delivery workers. We are talking about people who perform online tasks. So, they are, you know, facing a computer, and they perform small tasks like, I don’t know, they have to describe an image, or they have to transcribe a sentence. And all this is fragmented in small bits of informations, the information that we call data, and they are put in databases. And then, the models, AI models, use this information to learn. And this is why we talk about machine learning.

So, this is necessary, but because it is necessary, it has to be — this kind of work process has to be spread all over, literally, hundreds of millions of persons. And now the World Bank, in 2015, already estimated this type of workers in the tens of millions. And now, in 2023, they updated their estimate to hundreds. Now, they are very vague as to how many exactly. So, this is why they talk about something halfway between 100-and-something, 100-odds million and 400, which would represent something between 4% and 12% of the global workforce. Now, personally, I am a bit suspicious of these estimates. But one thing that I can say, given the fact that with our other colleagues, we’ve been trying ourselves to estimate these workers all around the world in several countries, we have noticed that over the years these estimates have been going up and up. So, we’re talking about probably 160 million persons in 2021, and now the estimates is bigger. And this is also something that we observe, because we are literally traveling the world, trying to, you know, give voice to this hidden workforce, and we notice that more and more persons are, well, available to speak with us and also organizing to protect their rights.

And you’re right, we’ve been to places like Madagascar and Kenya, and in some cases, we encounter people who are paid a few cents. This is the case of Venezuela, where at a moment the — even the factor of earning a few dollars, especially because it was in dollars, or in cryptocurrency, but not in bolívares, which is the local currency, which was depreciating from one day to the other — so, this was interesting, meaning this was an appealing type of job, let’s say. And in other countries, like, for instance, Madagascar, where we spent, well, a few months over several field works, we basically encountered at least one instance of a company that provided the opportunity for workers to connect to a platform, to work and to perform tasks on a platform, and they were paid in sugar, rice, beans and other foodstuff. So, when we say that they are paid peanuts, well, that’s a metaphor, clearly, but in this case, it was not a metaphor. It was something which was clear, visible, to the extent that they also had some kind of a meter at the very entry of this company to explain to the workers how much they would earn in terms of food for each type of — each level of performance, let’s say.

NERMEEN SHAIKH: Well, Professor Casilli, let’s listen to the data workers, precisely, in their own voice. Many data workers are tasked with reviewing highly disturbing content almost exclusively. This is former Kenyan data worker Faustine Makira in Kenya in a clip from the documentary In the Belly of AI. A warning to our listeners and viewers: It contains discussions of sexual abuse.

FAUSTINE MAKIRA: Well, at first, I didn’t see it as something that was hard. But as time went by, I realized that I used to have nightmares, especially when I’ve watched a lot of murder cases or rape cases, especially if it involves minors, children. When I’ve watched a lot of corpses during the day, I would have difficulty sleeping. I developed anxiety. I could not go out. I could not be where people were many in a crowd. I also felt like I was not being myself, because I was this — an outgoing person, but most of the time, because I was scared to go out, I would just be alone. And it kind of led me to depression, and that is why I decided to quit. Today, there are still traces of the PTSD or the things that I’m going through. That is why I’m still going for therapies. About my anxiety, most of the time, you will find me indoors. So, yeah, I would say there is still distress. Can we take a break, please? …

My co-workers, they did suffer the same symptoms. It just depended, like, with you. I have a colleague who is now a friend. He’s battling with insomnia. I have a friend who is also battling with anxiety. So, it just depends on how you, as a person, were affected, because you are affected differently. But, yes, we have very many — I have very many colleagues and friends who are also battling with PTSD in different ways.

INTERVIEWER: Why did not — didn’t you use the actual name of the clients and the companies you worked for?

FAUSTINE MAKIRA: Because they made us sign a nondisclosure form, which ties me to not say some things. Immediately I say them and they go out there, I will be jailed for more than 10 years. Yeah.

NERMEEN SHAIKH: So, that’s a clip from In the Belly of AI, directed by Henri Poulain. Professor Casilli, if you could just explain — you know, it’s not entirely clear — what do these data workers do with this bad content? You’ve called it something like “machine unlearning.” And then, we want to get also to the way in which these organizers against — these workers against impossible conditions are, in fact, trying to organize against their working conditions.

ANTONIO CASILLI: Sure. Well, I call this machine unlearning because sometimes these workers have the terrible task that consists in flagging information that they do not want the models, the AI, to reproduce, like they don’t want AI to create disturbing images or illegal content. They don’t want AI to perform actions that are considered, again, illegal or anti-social. So, they are somehow responsible, tasked for — with the difficult — the difficult task of performing this kind of teaching, which is a negative teaching.

And when we say moderation, content moderation in particular, the first thing we think about is social media, and we think that this is not connected to AI. Well, to the extent that AI exists on social media, too, and Meta has its own AI, and so on, TikTok has its own AI, you know, it is true that moderators can work both for social media and for AI. But it’s not only censorship. They are teaching the machine not to see some things, to remove them from the database. And this is something that is — that goes beyond the simple flagging of problematic content, because we have met people who perform this kind of machine unlearning also outside social media. I don’t know, somebody who has to train an AI that is used in surgery has to go through so many gruesome and bloody images, that can be traumatizing and can create the same type of PTSD, of post-traumatic stress disorder. Somebody who has to train, I don’t know, a self-driving car, they have to go through, for instance, car accidents, that can be extremely morbid and extremely problematic to watch itself, and they can develop this kind of traumatic mental health issues. So, it is something that is connected to the very process of teaching the machine to do something or not to do something, and this is also something that, you know, has a toll on these persons.

In this case — and I want to thank Faustine, that I have met again in Kenya a few months ago. I mean, for her testimony. But she is someone who paid an important, a serious, a severe personal price. Some other people were even less fortunate, because some other people died because of this, committed suicide or died in mysterious ways that are somehow connected to their activity. And this is something that we are documenting in many countries, in Kenya in particular, and so we are also trying to work with the families to, you know, face these kind of serious problems.

NERMEEN SHAIKH: So, Professor Casilli, could you explain: Why are these — why are workers made to sign nondisclosure agreements? And in some instances, as documented in the film, they aren’t even able to speak to one another about the content that they’re reviewing. Why is that?

ANTONIO CASILLI: Well, there are many reasons. The first one would be that they should not, according to the tech companies, talk about the kind of model that they are training, because this could constitute an issue in terms of intellectual property. For instance, if this company is developing a new model which is not on the market, they don’t want the competition to know that.

The second reason would be, yes, to actually invisibilize, as we say in a rather pompous way in academia, meaning to hide these workers. So, in order to hide them, the first thing is to, you know, well, shut them up, literally, and so that they — even if they are asked, they deny that they are actually working to produce AI, performing this kind of data work.

And one final reason is that, again, according to the designers of these AI models, if workers speak among themselves and they start discussing the kind of tasks they perform, this could introduce specific biases, meaning that they could, for instance, coordinate. This is the technical reason. But there is also a very — you know, reason that is connected to breaking the solidarity between workers. If workers are segmented, fragmented, they are separated from each other, and they do not speak to each other. They do not understand or they do not reach the awareness of the fact that they actually are doing the same kind of work: They are one class of workers. And this class consciousness does not develop.

NERMEEN SHAIKH: Professor Casilli, we’ll have to leave it there. Thank you for joining us, a professor at the Institut Polytechnique de Paris and author of Waiting for Robots: The Hired Hands of Automation. He also co-wrote the documentary In the Belly of AI, directed by Henri Poulain.

Coming up, Julián Posada, author of the forthcoming book Platform Extractivism: Data Work and the People Powering Artificial Intelligence. Stay with us.

[break]

NERMEEN SHAIKH: “Hi-Fly” by the late, legendary pianist and composer Randy Weston.