In Kenya, workers use clicks to train AI programs for major companies – under difficult and often traumatic circumstances.
Category
🗞
NewsTranscript
00:00Artificial intelligence is part of our lives, be it on social media platforms or in the
00:11AI chatbots we use. Among other things, AI makes sure we are protected from harmful,
00:16violent content. But there are real people behind this technology.
00:22We do not feel valued at all.
00:23Far from the world's glitzy tech hubs, these data workers train AI systems by sifting through
00:28massive amounts of harmful data.
00:32It tells you, explain to me, how does the flesh of a human taste like?
00:37While they carry out the often grueling task of keeping AI systems clean,
00:42international corporations on the other side of the globe reap the profits.
00:48Are we being used to make people billionaires?
00:52From Kenya and Germany, we set out to investigate this backstory of AI.
00:59Uncovering a web of global connections that power a hidden industry.
01:11Without us, there would be no end product because we are the most important part of this AI.
01:19What's life like for those training AI systems that the entire world now relies on?
01:28Our quest to find out begins in Kenya's bustling capital of Nairobi.
01:34Here we meet Joan Kinyua, a single mother of one.
01:38In her 20s, she left university to support her family,
01:42bouncing from one job to another until eight years ago, when she began taking on data annotation
01:47work that she found on platforms online. For years, this work mostly meant labeling
01:53elements in images.
01:54You annotate maybe vehicles, trees, people, like everything that is on the road that is visible,
02:00including road signs, but not in our roads, in the roads of San Francisco.
02:06Joan tacked all sorts of content. These odd jobs came in around the clock.
02:11She was paid by the task, often just cents.
02:16Over time, the work became more and more sinister.
02:20What does she mean by that? We'll get to that.
02:24But first, let's pause for a second to understand the role of data annotators like her
02:29in training cutting-edge AI systems. Most of today's AI analyzes huge amounts of data.
02:38The goal is to identify patterns, to identify patterns of data,
02:42and to identify patterns of data that are relevant to today's AI systems.
02:47The goal is to identify patterns. For example, how to identify a cat or a dog.
02:54This is how the system learns that cats tend to have pointed ears
02:59and triangular faces, while dogs tend to have broader faces and floppy ears.
03:09But what if the data isn't straightforward? This is where human data annotators come in.
03:17Their job is to make sure data makes sense to computers by teaching them.
03:21This is a cat. This is a dog.
03:27And these are neither.
03:34However, this often means dealing with questions far more dire than cat or dog.
03:40While pregnant, she sometimes worked 18 hours straight until 3 in the morning,
03:46pushing herself to the brink.
03:47While pregnant, she sometimes worked 18 hours straight
03:51until 3 in the morning, pushing herself to the brink.
04:08The World Bank estimates that hundreds of millions of people around the world
04:12are now engaged in online gig work, often with uncertain income streams and lack of contracts.
04:18In the West, many tech giants have started outsourcing this work to countries in the
04:22global south, such as Kenya. Data workers there are reported to sometimes earn less
04:27than $2 an hour, compared to more than $20 in the US.
04:34We meet another data worker, whom we'll call Faith. She asked us not to reveal her face.
04:41Faith trained large language models without being told which specific AI chatbot her efforts
04:46were supporting. With time, also her work became more troublesome.
05:05Faith tells us the job allowed her to help her elderly parents pay their medical bills.
05:10But there was only so much she could take.
05:25Faith realized this work was more than she could handle when she concluded her job was
05:30to train the chatbot to provide answers to user questions about these topics.
05:40But was she really training the chatbot to instruct people on cannibalism?
05:48What's behind all of this? We ask an expert to explain large language models.
06:01This is Aljosha Borkhardt, a leading expert in AI and large language models. We show him
06:07our interviews and ask him what he makes of Faith's account.
06:11For us humans, it's absolutely clear. I have an 11-year-old daughter. She knows nothing about
06:16cannibalism, I assume. To her, I could explain, I guess, in five to ten minutes what it is about.
06:22And from there on, she would know when it is okay to talk about it and what is out,
06:28and what is not okay to talk about. But for the machines, you have to give examples.
06:34Unlike his daughter, an AI system has no moral compass.
06:38That's why human data annotators like Faith are hired to add their tags.
06:43They help the AI understand when it's okay to talk about topics like cannibalism,
06:47that it's okay to provide historical information, for example, but not instructions on how to do it.
06:55It's all about context. There are contexts in which you want to talk about cannibalism,
07:00or when you talk about war situations, but you don't want to talk about cannibalism
07:05in the topic cluster of cooking recipes. And now this is the critical point where these people
07:14we are talking about today come into play. You would feed these examples to the chatbot
07:18as negative examples. That's why an AI chatbot we used gave us no instructions,
07:25and referred us to a Wikipedia article when we asked it to describe how to boil human flesh.
07:32Our research has shown that data workers like Faith often have little to no knowledge of the
07:37AI systems they're helping to train. Training them involves a complex network of companies.
07:44Faith says she found her freelance work through a Silicon Valley-based platform
07:48called Remotasks, a subsidiary of Scale AI. It's a company so successful that its CEO,
07:56Alexander Wang, briefly became the world's youngest self-made billionaire,
08:01and whose customers include U.S. tech giants and companies from around the world.
08:07Scale AI declined to be interviewed. In an email, a spokesperson wrote that
08:14training gen AI models to prevent harmful and abusive content is critical to the safe development
08:21of AI. Adding that we have numerous safeguards in place, including the ability to opt out at any
08:28time. The company we are working for is making millions or billions that makes this person
08:34really rich. Are we being used to make people billionaires? Like Faith, Joan spent years doing
08:42freelance work. One day in March of 2024, out of the blue, her employer, Remotasks,
08:48ceased operations in Kenya. These jobs, you do not have any security, especially
08:54in what you're working on. Like it's not even safe because how does a company where you've
09:00been working for more than five years just exit without any notice or anything? In an email to DW,
09:07a Remotasks spokesperson explained that people were not notified due to operational errors.
09:12So when I say that platforms and companies are predatory, that's exactly what I mean.
09:17This brings us to sociologist and computer scientist Milagros Miceli in Berlin.
09:24They go to a place as long as it is profitable for them, as long as it is convenient for them,
09:29and then they leave whenever they feel this is not great anymore. So what happens with Remotasks
09:36is just one case among many, many others that we have observed.
09:42In the summer of 2023, here at the Weizenbaum Research Institute,
09:48researchers launched an initiative called Data Workers Inquiry.
09:54We work with data workers so far in four continents.
09:59The situation on the ground depends on where these data workers live,
10:03but researchers say there does seem to be a common denominator. Workers were often marginalized.
10:10In my experience with doing this research in Kenya, one of the workers actually told me that
10:16they were told when they joined the organization that we employ the unemployable. And this was
10:20something that was constantly repeated so that you know your place. You come here, you do this
10:25work and you leave. So already you know, number one, you are unemployable, and number two, you're
10:30not here to grow. It is great that people who otherwise don't have a lot of chances get this
10:35type of jobs. The problem is not with that. The problem is with the quality of the jobs,
10:40their perspectives to actually grow. Although Remotasks has left Kenya, several other platforms
10:48continue to do business there. It's why Joan Kinyua has decided to join the country's growing
10:53data workers movement. On the one hand, it's about better pay. The union will advocate for
10:59fair working conditions, fair pay, and push for legislation and policies for people working in
11:05this field. On the other, it's about better psychological support. We work on very difficult
11:12projects, things like dead bodies and such, but there's not really, you know, it affects the mental
11:18state, but then nothing is provided for taking care of that, like paying for a psychologist as
11:23an example. Similar efforts by data workers to unionize suffered a setback last year when a
11:30local contractor working for US tech giant Meta fired 185 Facebook content moderators. Those
11:38dismissed are now suing the company. Meanwhile, Joan and her comrades in arms are continuing their
11:46fight for a union, and data workers in other countries do the same. Kenya, it's not a dumping
11:53place. Yes, we might have economical issues and crises, but we are not stupid. Why is it so hard
12:00for us to be recognized? Why is it so hard to be given for proper medical covers? Why is it
12:06hard for us to be maybe properly compensated and given good work conditions? If it is being done
12:13in other countries, then it should also be done in Kenya. As the AI revolution continues to ramp up,
12:20it's clear their fight is just beginning.