Feeding the machine: the benefits and costs for workers who power AI

In Venezuela, artificial intelligence work has offered an economic lifeline amid the country’s financial crisis. But a new study shows it comes at a cost.
Illustration: A data center with people working on artificial intelligence.

(AI-generated image)

For many workers in Venezuela, artificial intelligence has become an economic lifeline. Digital platform companies need workers to generate, curate, and verify the enormous amount of data that powers AI technology. And amid a protracted and severe economic crisis, the South American country supplies an eager labor pool.

But new research by Julian Posada, an assistant professor of American Studies in Yale’s Faculty of Arts and Sciences, finds that although these platforms provide a crucial wage-earning opportunity for Venezuelan workers, there is a costly power imbalance between employer and worker.

This disparity gives companies space “to dictate terms and leave workers vulnerable,” Posada concluded. “This erosion of worker autonomy, especially concerning financial compensation, underscores the overarching deterioration of work environments emblematic of the gig economy.”

The study appears in the April–June issue of the journal Big Data & Society.

Julian Posada
Julian Posada

Posada talked with Yale News about the vulnerability of data workers, the economics of “deep embeddedness,” and the need to expand the ethical debate over AI to include its production. The interview has been edited and condensed. 

How did Venezuela become a hub for these data production platforms?

Julian Posada: Venezuela had a socialist government under [its former president] Hugo Chavez. He invested a lot of the oil revenue from the country into infrastructure and welfare programs and so on. But he badly mismanaged the economy. Venezuela was completely dependent on oil exports and there was also corruption, political instability, and foreign sanctions. In 2014, oil prices dropped dramatically, causing an economic crisis that continues today. Venezuela had the worst hyperinflation on the planet for many years and inflation is still very high.

The country became a hub for these platforms for three main reasons. First, the country had the infrastructure in place — electricity and internet connections. Second, the economic crisis and hyperinflation caused the local currency, the bolivar, to be highly volatile. And third, unemployment was high. People started looking for online work and work that paid in U.S. dollars, as the dollar became the de facto currency in the country. When the AI boom happened, Silicon Valley needed a lot of data and workers to curate that data. They started to outsource that work to these platforms. And Venezuela provided a cheap, easily exploitable labor force.

What do these workers do?

Posada: Some of them generate data. For example, take photos of themselves, then feed the images to machine-learning systems to create facial recognition algorithms. Some annotate data or give meaning to data. For example, with a bunch of images of a kitchen, the worker will label the objects — this is a window, this is a light, these are paintings, and so on.

Workers also doing a lot of verification. Machines learn through iterations on their own, and humans need to tell the machine, “You’re doing well” or, “You’re doing bad.” And very recently workers have started “red teaming,” meaning they’re trying to exploit the system to find vulnerabilities. They’re trying to make the algorithm produce things like illegal content or exploitative content so the engineers can prevent it from doing that.

Your paper explores “deep embeddedness” and how it impacts these data workers. What do you mean by deep embeddedness?

Posada: Embeddedness in economic sociology means that economic behavior and institutions are deeply entwined with social relationships and structures. For example, when you buy eggs at the supermarket, you are embedded within a network of workers, distributors, farmers, laws, and so on. There’s a network of people and institutions that make economic activity possible. Deep embeddedness basically means that there are more actors involved in those economic transactions.

In the case of the data workers, they have their employer — the platform — and the employer connects them with the Silicon Valley company. At the same time, the platform doesn’t pay them directly. They use PayPal, or something like it. The workers are also connected to electronic wallets, where they can transfer their pay. Normally, the more people involved in a network, the better for the actors. However, my study finds that in the case of freelancers, the more actors, platforms, institutions involved in these networks, the less beneficial it is for the worker in the end, in terms of income.

You found that these platforms do offer benefits to the workers in that they give them access to U.S. dollars, which, as a more stable currency, enables them to save some money. But there are significant financial disadvantages as well. What are those?

Posada: The first one is loss of autonomy. The platforms pay through platforms like PayPal, and workers have to accept the fees they charge — they don’t have any say. The second one is decreased wages. Platforms like PayPal charge a commission, and then workers move money to an electronic wallet, which charges a commission as well. Then if workers exchange dollars for bolivars, the brokers charge another commission. So there is a loss of income throughout the chain.

Much of the talk around ethics in AI has focused on the quality of the data being generated — whether it’s biased, for example — but far less attention has been paid to the treatment of data workers. Do you see that changing?

Posada: Not yet. At first, we talked about ethics principles, now we’re talking about governance creating laws. We have the executive order on safe, secure, and trustworthy AI from the Biden administration. The European Union just passed the AI Act.

The problem is we still have a consequentialist approach to AI, meaning that these efforts are geared toward the deployment of the technology. Not about how it comes to be, about its development. We don’t talk about environmental costs; we don’t talk about human capital or exploitation of workers. That’s the fight for the next decade as it keeps growing.

Share this with Facebook Share this with X Share this with LinkedIn Share this with Email Print this