You may have heard that revolutionary AI sits on old-world foundations. The supply chain churning out generative AI tools like ChatGPT has highly paid executives and researchers at the top, and at the bottom, working stiffs who toil at screens training algorithms. Between 150 million and 430 million people do such work, according to a recent World Bank estimate: They annotate images, text and audio; create bounding boxes around objects in images; and, more recently, write haikus, essays and fictional stories to train the sophisticated tools that could eventually replace people like me.
They also exist in a kind of economic stasis. “I’ve never met a worker who would tell me, ‘This job gave me the chance to buy my house or send my kids to university,’” says Milagros Miceli, a researcher at the Distributed AI Research Institute and Weizenbaum Institute who has worked with scores of data workers across the world.
Miceli recalls speaking to about a dozen data-labeling workers earning about $1.70 an hour in an Argentina slum in 2019. When she returned in 2021, none had moved on and their wages had barely increased. They were still living below the poverty line.
Workers often have to take second jobs or night shifts, says Madhumita Murgia, the AI editor of the Financial Times whose recent book “Code Dependent” features their stories from across the developing world. One woman who worked for Samasource Impact Sourcing in Nairobi, Kenya, for instance, couldn’t support herself and her daughter on her salary and had to move in with her parents, Murgia says.
The job itself is precarious. Another worker in Bulgaria couldn’t make rent because she was suspended from accepting paid tasks after complaining about night shifts. “You’re one step away from everything unraveling,” says Murgia. End customers are the likes of Microsoft and OpenAI, some of the most valuable firms in the world. “It’s like the factory worker in the Philippines who doesn’t realize the dress they’re stitching is going to be a $3,000 gown.”
There is also precious little of that time-honored aspiration for the developing world: upward mobility. Murgia found that data workers weren’t transitioning to higher-paying digital jobs. “They’re still confined to low-value work,” she says.
Leaders of data-labeling firms often start with noble intentions to help pull people out of poverty, but they’ve struggled to get corporate customers to pay higher rates as competition in their field has increased. As such, most data work platforms don’t have policies in place to ensure that their workers earn at least the local minimum wage, according to a 2021 survey from the Oxford Internet Institute.
Take this job ad seeking “professional translators” in Nigeria that offers up to $17 an hour to help train generative AI models. That is well below the average rate for Nigerian translators, who tend to start at $25 an hour, according to Good Firms, a client-reviews website. The ad comes from Remotasks, the main platform of San Francisco-based AI startup Scale.ai, which just raised $1 billion from investors including Amazon in one of the year’s largest financing rounds. Scale.ai didn’t respond to multiple requests for comment.
The company and rivals like San Francisco-based Samasource Impact Sourcing, Argentina’s Arbusta and Bulgaria’s Humans in the Loop play a critical role in the AI supply chain, but for years now have typically paid just enough for workers to maintain a living, Murgia and Miceli say.
That may continue even as data work becomes more complex. Recently, platforms like Scale.ai have been looking for more skilled workers, including artists and people with creative-writing degrees to write short stories for training AI systems, according to instruction documents seen by Miceli. While those offer higher wages, they are still below what people with degrees should be earning.
Researchers say the appetite for such work is growing, but with few incentives to provide an equitable wage, it’s hard to see workers’ economic status improving. Training AI is already horrifically expensive due to the cost of chips and cloud computing. (Venture capital firm Sequoia Capital recently calculated that the AI industry spent $50 billion on Nvidia chips to train AI in 2023 but only made about $3 billion in revenue.)
That spells fewer opportunities for the people underpinning the AI revolution and shows yet again that the technology’s true transformative effects have been in entrenching economic power.
Perhaps we can learn something from Nike. Back in the 1990s, the company faced an enormous backlash for the long hours and meager wages its workers in developing nations earned. Over time, consumer boycotts and pressure from the media led Nike to put in stricter labor policies. It spent millions of dollars on improving conditions and pay.
The challenge for data workers is that their jobs are harder to visualize in the same, concrete way you can imagine a young boy sewing tennis shoes in a dimly lit warehouse, and that can make it harder for their advocates to rally support. But tech companies should remember that poor working conditions at the bottom of their supply chain can also lead to substandard AI. That’s problematic at a time when the public is more wary than ever of buzzy models that hallucinate. The answer to that isn’t rocket science: Pay the data workers more, and treat them better, too.