Under the sleek, polished glass of earlier models of Apple’s iPhone, you might have found cobalt—the blue-stained mineral nicknamed the “blood diamond of batteries”—mined by child labor in the Democratic Republic of Congo. After a report on unethical mining by Amnesty International in 2016 and a Sky News dispatch in 2017, Apple suspended use of hand-mined cobalt. But, according to reporting in the New Yorker in 2021, “once the media attention died down the practice continued,” and the linked company, Zhejiang Huayou Cobalt, remained part of Apple’s supply chain. Apple is far from the only tech firm implicated in such practices. In many cases, a technology’s shiny exterior hides exploitation beneath.
The boom in artificial intelligence (AI) and large language models over recent years have not been exempt from this trend. On Monday, the Washington Post published an investigation into the “digital sweatshops” of the Philippines where workers train AI models for often below minimum wage pay. In interviews with thirty-six current and former freelance workers for the company Scale AI, which has delivered services for Meta, Microsoft, OpenAI and the US Department of Defense, the Post found payments routinely delayed or canceled. “The budget for all this, I know it’s big… None of that is trickling down to us,” Paul, one of the workers, told them. As the Post wrote, “While AI is often thought of as human-free machine learning, the technology actually relies on the labor-intensive efforts of a workforce spread across much of the Global South and often subject to exploitation.”
One reporter who has done much to uncover these exploitative relationships is Billy Perrigo, a correspondent at TIME based in the London bureau. His investigation exposing “Facebook’s African Sweatshop” led to an ongoing lawsuit against Facebook in Kenya and a successful vote to unionize by impacted workers; his piece on how “OpenAI Used Kenyan Workers on Less Than $2 Per Hour to Make ChatGPT Less Toxic” energized the conversation on AI’s unscrupulous practices. At its best, Perrigo’s reporting on exploitation by big tech firms, which was shortlisted for last year’s Orwell Prize, is like pulling up a marble paving stone in Silicon Valley to find the dirty practices hiding beneath. Our conversation has been edited for clarity and length.
JB: Earlier this year, you reported on how OpenAI relies on Kenyan workers paid between $1.32 and $2 an hour to make ChatGPT less toxic, trawling through stacks of traumatic data to train its algorithms and detect problematic content. Can you tell us more about the story and how you reported it?
BP: The crucial thing about this story was that it wasn’t actually the first story that I wrote about the outsourcing company in question [Sama]. I’d published a story before about how the same outsourcer had done content moderation work for Facebook in quite similar circumstances. Basically, low-paid employees having to look at the worst material that you can imagine, so that the rest of us don’t have to see it. After the publication of that story, I became aware that OpenAI had also been a client.
This was before the mass excitement about large language models, but I’d been playing with GPT-3 for a long time and sensed that kind of boom was coming. As I was reporting the story out, OpenAI released ChatGPT, and suddenly the entire world became aware of this technology that Sama, the outsourcing company, had been helping OpenAI to build. Effectively, employees would read snippets of text and categorize them for different types of toxicity, including violence and sexual abuse. Because OpenAI, for very good reason, didn’t want their chatbot to be sexually abusive or hateful.
What did some of the employees doing this work tell you about the impact on them?
They all said the work was basically quite traumatic. And it took me a while on the reporting process to fully internalize how traumatic that work could be. I’d been used to talking to Facebook content moderators who are viewing images and videos, where it’s much easier to understand how something could be causing vicarious trauma. But with reading text, it wasn’t immediately clear to me how that would cause trauma. One day, one of my sources said to me, I’m paraphrasing, “You’re reading this content, day in, day out, over the course of days and weeks and months, it seeps into your brain and you can’t get rid of it.” I saw many examples of the kinds of texts that they were reading, and they were pretty grim. There’s been subsequent reporting on this from some very good journalists, which elaborates in a lot more detail than my initial story about the personal ramifications of this work on some of the people who did it, that involves marital breakdown, depression.
You alluded to your previous story on how Facebook is using the same company, Sama, to train its algorithms to detect horrific content. And your investigation found workers—paid as little as $1.50 an hour—becoming traumatized, and the company crushing a unionization drive. Can you talk about how tech firms rely on these massive supply chains of human labor, which is often missing from the conversation?
A lot of the Silicon Valley narrative is that the technology these companies are building is shiny and clean and divorced from any of the material difficulties that you might see in the rest of the world. Looking at supply chains, whether that’s supply chain of data or raw materials, that picture quickly falls apart. In the case of social media, we’ve known for a long time—and I was by no means the first to report—that the work of keeping feeds free of the most grim material cannot be done by AI, and largely relies upon human labor. Even as these companies are using AI to reduce the number of humans who are doing it, there is no such thing as an AI tool that can do this work to the level that a human can.
It’s an open question now whether that is changing. OpenAI recently put out a research paper that said GPT-4 can actually do content moderation to roughly the same level as a human. If it’s borne out, that’s actually quite good news in the sense that it doesn’t mean content moderation going forward will rely on this mass of human labor. But then you ask the question, who does that technological innovation benefit? And it doesn’t actually benefit any of the people who were doing that content moderation in the past, and now they will lose their jobs. The work was sometimes talked about in the language of providing opportunities for people in the Global South. If that work is to be automated away, it’s unclear to what extent those people will benefit.
Is this a new form of colonialism in a way—big tech platforms, in your words, “outsourcing trauma to the developing world”?
I wouldn’t say it’s necessarily a new form of colonialism, but it certainly benefits from historical inequities of power and extraction. It does more to concentrate wealth and power than it does to redistribute.
In your reporting on big tech from below, you speak to people who are not only suffering with PTSD or psychological trauma, but also their livelihoods are at stake if they’re discovered speaking out. As a journalist, how do you navigate that with sensitivity?
For the Facebook story, we worked with a whistleblower protection group called the Signals Network, who were great. They connected Daniel [Motaung], the whistleblower, with a therapist, and that’s where he was diagnosed with PTSD, he hadn’t actually received a formal diagnosis before. But it was very clear from simply interacting with him that he was not in a good place. We gave anonymity to all of the other sources in that story who requested it. Because, you’re correct, there are significant repercussions for speaking out. There’s the widespread use of non-disclosure agreements, which often actually don’t necessarily have as much of a legal basis as they’d have us believe. But regardless, when you are as far down on the wealth and power scale as many of these people are compared to their employers, that’s a very real consideration.
But on the other hand, many of these people have so little to lose because of the position they’ve been put in, that at least some of them that I spoke to felt they had no other option but to speak up.
Since Elon Musk took over Twitter last year, the company has rowed back on content moderation, with studies showing that hate speech has jumped on the platform since then. And this has given cover to others like Facebook to bring back formerly banned accounts too. What’s your view on the impact of this weakening of moderation recently?
It’s clear that it’s coming at a time when we need it perhaps more than ever. Not only do we have a very contentious presidential election coming up in the US, but also the rise of artificial intelligence tools, which can produce not only disinformation but also targeted harassment and deep fakes and abuse at scale. We’re living in an increasingly fractured media environment, where even if it wasn’t for the rise of AI tools that completely flub the truth in a significant percentage of their answers, it’s becoming more economically difficult for news organizations to even survive. So content moderation isn’t a little thing on the side that’s unrelated to all of those things. It’s pretty central to questions of, what information environment are we living in? The decline of tech companies investing in that as a priority should be seen holistically with all of these things.
To their credit, OpenAI and the other large language model purveyors have learned many of the lessons that social media companies didn’t take heed off in the first ten years of their lives. OpenAI was right to try and build those protections into the chatbot. It’s just a shame that the work was handled at that outsourcer the way that it was. Also, Sama is simply the only outsourcer that we know of that was doing this work for ChatGPT.
The 2024 election cycle is beginning to gear up now. What do you expect from what could be the first AI election?
Yes, great question. I’m not an expert by any stretch. One thing that several AI CEOs have said now, including [OpenAI’s] Sam Altman, is that although these tools can be used to generate disinformation at scale, they can’t be used to distribute disinformation at scale. To distribute information as well you need a megaphone, like a social media platform or a traditional media platform. I haven’t seen good evidence yet that large language models are making the distribution of disinformation and other kinds of information easier. That’s not to say it isn’t changing the character of disinformation and the difficulty of dealing with it. I think they degrade the information ecosystem in many other ways. It’s going to be pretty tough. But we shouldn’t forget about social media as the vector of spreading these kinds of various harmful types of content.
TIME has interviewed industry leaders in AI, from DeepMind’s Demis Hassabis to OpenAI’s Sam Altman. Your [co-authored] piece characterized Altman as an “evangelist preaching about risks yet plowing ahead anyway.” I wonder what impression you get from the people designing these technologies of their awareness of their social responsibilities?
I didn’t actually interview Altman, that was my colleague Edward Felsenthal. I get the impression from both of them that they are very aware of the social impact of their work, and they are trying to mitigate the harms as much as they can. The more pertinent question is: what incentives are they bound by? And what structures have they set up for themselves in order to prevent the harm and maximize the good of these systems? That’s a major unanswered question that will define whether AI progress is good or bad. I don’t think most people at Facebook are acting maliciously and I don’t think anyone at OpenAI is acting maliciously. But we live in a society where incentives matter more than intent in many cases.
