Профессор Йельского университета описал эксплуатацию венесуэльских работников данных для обучения искусственного интеллекта

Julián Posada’s upcoming book, titled Platform Extractivism: Data Work and the People Powering Artificial Intelligence, is based on extensive fieldwork among data workers in Venezuela. During the height of the COVID-19 pandemic, amid a hyperinflation crisis, entire communities in Venezuela depended on data work, says Posada, a professor at Yale University. Paid as little as 50 cents per hour and monitored by surveillance algorithms, most workers did not even know who or for what end they were working for. “From the perspective of the worker, you would just see the tasks. … [T]he algorithm here is the manager.” Transcript NERMEEN SHAIKH: We continue our discussion of the hidden data workforce powering artificial intelligence with Julián Posada, author of the forthcoming book Platform Extractivism: Data Work and the People Powering Artificial Intelligence, which is a deep look at data work and data workers in Venezuela. Julián Posada is assistant professor of American studies at Yale University and co-director of its Computing, Culture & Society certificate program. Julián, welcome to Democracy Now! If you could — we just heard from professor Antonio Casilli about data workers. You focused principally on Venezuela. If you could tell us what you found among data workers there? JULIÁN POSADA: Yes, so, when I started this project, I actually wanted to focus on Latin America, where I grew up. And the first thing that I found when I started this research is that most of the workers, at the time when I started, around 2020, were located specifically in Venezuela. So, part of the project and the book is about why Venezuela contains so many of the workers that were powering — what I call in the book, “powering artificial intelligence” at the time. I found three reasons for that. The first reason is, of course, the economic conditions of the country. Since 2014, the economy of Venezuela has been in a crisis — hyperinflation, high unemployment rates. And adding to that, you have then the COVID pandemic starting, increasing unemployment rates with the lockdowns. So, it was really, really a dire situation in the country around seven, six years ago. The second reason, however, why Venezuela became so prominent in data work is the infrastructure of the country, which was developed by the government. So, for example, people have access to computers that were created in Venezuela, built in Venezuelan factories during the Hugo Chávez years, that then were retained by families. And those same computers were the ones used for data work during the pandemic. And then, the third reason, which is also an interesting one, is the composition of families in Venezuela. What I found is that, from the qualitative interviews, not a single worker was alone, workers really depending on their families. And during the pandemic, when a lot of people lost their employment, it was just one breadwinner in the house, the data worker, and the entire household was supporting that work through many activities — feeding the worker, doing house chores and so on. But the family really revolved around data work, and then the neighbors, as well. If you lost access to electricity, if you lost access to something, your neighbors would be there to provide support. So, this is why I think Venezuela became such a prominent case for data work in Latin America and in the world during the pandemic: again, the economic crisis, the infrastructure and the composition of families. NERMEEN SHAIKH: And could you tell us how — what were the salaries like for these workers? What were the working conditions like? And who were they working for? JULIÁN POSADA: So, salaries were either very low, some platforms — a platform, Remotasks, would pay workers per hour, and it would be 50 cents of a dollar per hour. Other platforms, the majority of them, would pay per task, so it would be a form of piecework. Workers would earn around $2 to $5 per week because of the hyperinflation then. Those few dollars were important for their income, right? And the interesting part is that some workers — and this is tied to the question about the number of workers and their location — some workers would use VPNs to mask their IP addresses, pretend they’re in countries like here in the United States, and then get paid more for the same amount of work, pretending not to be in Venezuela. So, it’s really difficult to, one, estimate who they are, where they are, but also their pay rate really varies across tasks and platforms. NERMEEN SHAIKH: And how do we — do they have any idea for what major tech company they’re working? Like, does OpenAI or Anthropic — do they — how do they employ these workers, effectively, through — how does it work? JULIÁN POSADA: So, a company would usually approach an outsourcing company. Let’s say, for example, you have your AI startup. You approach one of the outsourcing companies, for example, Scale AI, which is the company behind Remotasks. Then, with that company, you would then create a project and then outsource it to one of their platforms — in this case, Remotasks. And they would have programs — in this case, for Venezuelan workers. Then, from the perspective of the worker, you would just see the tasks. In most cases, you don’t know who they’re working for, what task you’re doing for, or what the AI that you’re doing is in reality. I asked some workers if they had any guesses. For example, one worker was telling me, “I think I worked for a task related to the military complex, because I have to tag roofs and bridges in some deserted place somewhere. But I have no idea where this is, and I have no idea what the company or what the AI I’m training is actually for.” So, there was really complete invisibility in that regard. The workers would not know who their employers were and what even the AI they were helping develop is. NERMEEN SHAIKH: And if you could talk a little bit also about how the workers are monitored and how their work is quantified? JULIÁN POSADA: Yeah, so, like in any case of gig work, the algorithm here is the manager. And there are some cases in which workers, like in the case of Kenyan workers or workers in Madagascar, for example, where they’re working for companies inside. They go to an office, like a call center. They’re working in front of a computer. They have their boss around. And they go to a single location. That’s what we in the academia call the business processing outsourcing centers, BPOs. My work in Venezuela was primarily focused on platforms. So, the best way to think about it is think of the Uber of data annotation. So, you would log in to your computer, to a platform, on your browser, and then you would access the tasks and then do the tasks. And in this case, the algorithms will be the ones, the managerial algorithms, the ones actually controlling the workflow. You would have a timer. They will know how long you’ve been in each task. They will know — try to guess the accuracy of your tasks. They would present the same task twice or three times, so if you’re doing different things, the algorithm will think, “You’re spamming, so I’m going to ban you from this task now.” So, really, it was an algorithm put in place to control the workforce and ensure that they were doing the job that the clients wanted them to do. NERMEEN SHAIKH: And what was the thing — were you able to speak to a lot of these workers? JULIÁN POSADA: Yes, I did. NERMEEN SHAIKH: And what was the most frequently reported thing that they told you? JULIÁN POSADA: Well, one interesting thing is that, really depending on the task they were doing, they really varied in terms of payment, because of the investment. So, for example, there was this task called “Beast Mode,” in which workers would have to label images in households. Let’s say an image of a living room, you have to label the couch as a couch, a mirror as a mirror, and so on. And workers would then be paid bonus for these tasks that were really important for the companies and the firms. In these tasks, they were very repetitive, most of the cases. And at the time — so, this is pre-LLMs, right? So, at the time, there was a lot of investments on computer vision, investments, for example, to develop facial recognition algorithms or object identification algorithms. So, this is why Venezuela, in which the workers — many of them didn’t have knowledge of the English language — were able to tag these objects and train, then, these image recognition algorithms. Then, post-2023, when the investments started to focus on large language models — this is when ChatGPT came out — then a lot of that work was moved from Venezuela to places with, I would say, a colonial legacy with English-speaking countries, so the Philippines or India, in which many of these workers were required to speak English or knowing the English language, then were tasked with any tasks that would do with language, the English language in this case. And at the same time, the situation in Venezuela improved a little bit. Oil prices started to increase again. And a lot of these workers then stopped doing data work and went back to their regular activities. NERMEEN SHAIKH: Thank you very much, Professor Julián Posada, assistant professor of American studies and co-director of the certificate in Computing, Culture & Society at Yale University. He’s the author of the forthcoming book, Platform Extractivism: Data Work and the People Powering Artificial Intelligence. And that does it for today’s show. Amy Goodman will be in Middlebury, Vermont, Friday night for a screening of the film about Democracy Now! called Steal This Story, Please! with the Oscar-nominated directors, then on to Madison, Wisconsin, on Labor Day weekend. Check our website at democracynow.org for details. Democracy Now! is produced with Mike Burke, Deena Guzder, Anjali Kamat, Messiah Rhodes, Nicole Salazar, María Taracena, John Hamilton, Sara Nasser, Charina Nadura, Sam Alcoff, Tey-Marie Astudillo, Robby Karran, Hany Massoud and Diego Ramos. Our executive director is Julie Crosby. And very special thanks to Becca Staley, Jon Randolph, Paul Powell, Mike Di Filippo, Miguel Nogueira, Hugh Gran, Carl Marxer, Denis Moynihan, David Prude, Dennis McCormick, Matt Ealy, Anna Özbek, Emily Andersen, Dante Torrieri and Buffy Saint Marie Hernandez. I’m Nermeen Shaikh. Thanks so much for joining us for another edition of Democracy Now!