Investigating human-powered AI

PhD student Julian Posada’s research into the human workers behind the Artificial Intelligence industry has landed him a tenure track position at Yale 

Before the pandemic struck, Julian Posada planned to travel to Venezuela to interview the people who do outsourced artificial intelligence work for foreign companies. But disappointed as he was to have to cancel those plans, the Faculty of Information student was still able to forge ahead with the research for his doctoral dissertation.  

The digital workers Posada needed to interview used Zoom for their work and were eager to share their experiences online. “The pandemic didn’t stop them from working nor from granting interviews,” he said. In fact, thousands more joined the online data workforce due to Covid-19. 

These data workers, as Posada and fellow researchers define them, are involved in the collection, curation, classification, labelling and verification of data. They are employed, on a piecemeal basis by what are known as AI platform companies, who outsource their data work to countries where labour is cheaper. 

Posada’s research, which focused on Venezuelan data workers, also examines how power imbalances and coloniality are reproduced in this particular subset of the gig economy. He argues that data quality and AI algorithms would both be improved by allowing workers more input and offering better labour conditions. 

To protect the workers he interviewed, Posada does not name their employers in his research, but he explains that these AI platform companies have clients like Uber and Google as well as smaller start-ups that no one has heard of. The workers are assigned to work with the data needed to operate everything from voice assistants and chatbots at airline and telecom companies to self-driving cars. 

A day’s work could include data generation tasks like scanning items at the grocery store, submitting photos of oneself as an adult and a child to create a process for aging, or photographing and uploading pictures of people in different positions to help machines learn about human body movements.  

More common tasks include data annotation, which can involve labelling and classifying objects in street view and other videos, and verification, where humans evaluate whether AI is making correct assessments. This is done for everything from search engine results to the quality of chatbot conversations.  

Posada focused his research on Venezuela because it is the second largest source of data workers after the United States. Before the country’s oil economy collapsed, there had been major investments in building online infrastructure including providing homemade computers to school children. The collapse caused many people to lose their employment creating what Posada calls “the perfect characteristics for outsourcing jobs.” 

As a Spanish-speaking Colombian, Posada could talk easily to the data workers interviewed for his thesis and he also made the decision to compensate them for their time. The dozens of people interviewed included a retired couple unable to make do on their pensions alone, an IT specialist in need of a second source of income, and a family of five – mother, father and three children – all working for a platform company.  

While the data workers are grateful to earn a few US dollars per week, they are constantly aware of the fragility of their situation. In social media groups, says Posada, workers counsel each other not to take on tasks with bugs as they will be penalized. Reporting the bugs, they realize, is too time consuming and will result in a smaller paycheck. 

In his doctoral work, Posada worked under the supervision of Associate Professor Alessandro Delfanti, a digital labour expert. He was also strongly influenced by the AI research carried out at UofT’s Centre for Ethics and the Schwartz Reisman Institute for Technology and Society, where he was a fellow in his final year. While he arrived in Toronto from France, where he completed both his undergraduate degree and a Master’s in sociology, with a good knowledge of how platform companies operate in the gig economy, it was not until he began his doctoral studies that he realized how well established the artificial intelligence industry already was. 

“I was upset people were not thinking about labour when they were thinking about AI,” says Posada. In a 2020 commentary he wrote, he argued that current discussions on AI and labour focus on the deployment of these technologies in the workplace but ignore the essential role of human labour in their development. 

As a result, Posada shifted the focus of his planned thesis to look at the topic through a data production lens, a change that he believes eventually led Yale University to offer him a tenure track assistant professor position in American Studies. “My current research on how data production is outsourced can be expanded to include other ways to study data from multiple fields,” he says adding that data science should not only be a methodological field but also have data as an object of study. 

Posada will take his assistant professor role in July 2023 after a year of postdoc work, also at Yale. He is currently preparing a book proposal based on his doctoral work and tentatively titled, “Extract and Impose: How Artificial Intelligence is Perpetuating Coloniality.” 

As he continues to explore the world of outsourced digital work, Posada is turning his attention to robots and robotic devices piloted by overseas workers. “It’s another step in the evolution of outsourcing,” he says giving the example of a company in California using delivery robots piloted by people in Colombia. “What does it mean for future migration when this virtual migration can happen?” 

Looking at the bigger picture, Posada hopes that the fields traditionally associated with data science, like computer science and statistics, will see how they can benefit from the qualitative research approaches used in the humanities and research. “Data science techniques have already influenced the humanities and the social sciences, spurring innovative sub-fields like digital humanities and computational social sciences,” he says. “I think the opposite should occur for computing and statistics.”