What It’s Like to Be a Data Labeler Training AI

0:03Hello, and welcome to the 404 Media Podcast, where we bring you unparalleled access to hidden worlds both online and IRL. 404 Media is a journalist founded and owned company and needs your support. To find our work and subscribe, go to 404media.co. Subscribers get access to bonus podcast segments and early access to interview episodes like this one. I'm Jason Kevler. And this week, we have a special episode and something a little bit different. I recently went to Kenya for a journalism and AI conference that I talked about briefly on this pod before. And while I was there, I really wanted to meet with Michael Jeffrey Asia, who is the Secretary General of the Data Labelers Association.

0:43Data labeling is a huge job in Kenya. A lot of people there are talking about it all the time. It's seen as tech work because it is tech work. And for people who don't know, data labelers are the people who train AI and who also work on ensuring the outputs are accurate. In some cases, data labelers are themselves pretending to be AI in order to eventually train AI tools. A lot of times data labelers don't know exactly what they're working on or who they're working for because the work usually goes through a platform or a subcontractor or a combination of both. Although, as you'll see in this conversation, a lot of data labelers are able to figure out who exactly they're working for.

1:24But basically, they can be presented with a backend where they're asked to perform tasks or correct outputs or answer questions or label like video, photos, things like that. And in some cases, as you'll see, their answers are actually presented in real time as AI to the end user. And so there's been all these stories where in quote unquote, like AI messenger is actually just some data labelers responding and pretending to be AI, even though the data labelers themselves might not know that that's happening. Data labeling is notoriously brutal and underpaid work. Workers sometimes earn as little as a few dollars per day.

2:05They work under algorithmic management where they have these really strict quotas. And then because they're sometimes trying to train AI about what not to do or what not to show, they're often shown graphic, violent, or sexual content for hours at a time. This has led to a lot of cases of PTSD. There's actually a lawsuit, a couple lawsuits in Kenya going on right now about this. Meta has been sued. There's a company called Sama that Michael worked for that has had a lot of complaints against it. And data labeling is really similar to content moderation jobs. And a lot of people who work in data labeling also do content moderation or they switch back and forth between the industries.

2:47It's such a big thing in Kenya at the moment that sort of driving down the highway, going into Nairobi, you see all of these huge office complexes where people do data labeling and like big SEMA offices and things like that. And I actually was like, I mentioned data labeling to the driver who took me to meet Michael for this interview. And she told me that she was also a data labeler and so are a lot of her friends. I wanted to talk to Michael because he's the author of a report called The Emotional Labor Behind AI Intimacy, which was put out a few months ago by the Data Workers Inquiry Project.

3:20In this report, Michael explains working endless hours pretending to be an AI sex bot or a bunch of different AI sex bots, more or less, and the toll that that took on his mental health and his marriage. He's going to talk about that in the interview, but I want to read a passage from it that was really affecting for me. It goes, quote, I had to assume fabricated identities, memorizing false backstories, and reading through previous chats. Sometimes I would be assigned a conversation that had been ongoing for several days and had to continue it smoothly so the user wouldn't realize the person responding had changed.

3:55I played the part, stepping into carefully crafted personas designed to connect with unsuspecting customers on a quote-unquote personal level, often through sexual or intimate conversations. When I logged into my work dashboard, I had access to multiple fake profiles of varying genders, typically three to five different personas I could operate simultaneously. Sometimes I had to operate male and female personas on the same day, depending on what the platform's users were seeking. One day I might be Jessica, a 24-year-old lesbian college student from California, and Joe, a 30-year-old gay man from Florida.

4:30Another day it could be Maria, a 28-year-old heterosexual nurse or a nameless woman artist. I felt like I was losing myself in the role. It started as any other job, responding with empathy and willfully pretending to care. But over time, it became harder to separate the act from reality. The lines blurred. I began questioning if I was acting or if I was truly becoming the persona I was forced to embody. I was losing touch with who I really was, a feeling that has never left me. Michael has since become really critical at the Data Labelers Association, a group that is fighting to organize people who do data labeling work in Kenya, and I guess internationally, actually, and who's advocating for better working conditions, higher pay, and more protection for data labelers.

5:12I met Michael at a co-working space in Nairobi in a very tiny room. So I'm not on camera after this, but here's my conversation with Michael.

5:27I guess first, can you introduce yourself? My name is Michael. Currently, I'm the General Secretary of the Data Label Association, having built this space for over five years now. So I'll basically take you through what data labeling is. And I'll use a practical example of a self-driving car, for example. If you have to have this car on the road working without causing accidents on the roads, this vehicle has to be taught a lot of things. It has to be fed with a lot of information, for example. If this vehicle has to identify, let me say, a truck, an SUV, a person, a door, you know, any other animal, you know, the cabs, the whole springs behind the sides, the roads, the traffic signs, the traffic lights, and what have you.

6:09this vehicle has to be taught. It has to be given this information like a small child. So when you have to help it understand that it's a human being, you don't take a picture of a human being, just one person. You know, you have to take millions and millions of pictures of people in different heights, different sizes, you know. So that by the time you're working on this, you're giving it a variety. You know what people are, because if you gave it, let me say pictures of grown-ups only. If it finds a kid on the road, it's going to hit the kid. So all of these pictures are taken on the roads randomly.

6:38then you're given this information. So you're told to label these vehicles correctly. And that is why anytime you're doing all this labeling, you have to be very accurate. You have to identify this as a bus. You have to identify this as an SUV, for example. This as a truck. This is a trailer. So that it doesn't hit any of this. It doesn't cause an accident. So that is why when this information is brought to us, we have to label it up correctly. Once all this is done, the information is put together, then this machine is stored. You know, through an algorithm. on how to identify all these things.

7:11Yeah, so I know that there's data labeling like businesses kind of all over the world, but I know that there's a lot of it in Kenya. Do you know why it's such a popular job here? First of all, Kenya has, I think one of the best internet penetrations worldwide, if not within the continent of Africa. I think almost every homestead in Kenya has an internet connection at the moment. Second, Kenyans are learned. most Kenyans have gone through a very good education system, and most of them are aggressive. I think Kenyans, I call it a challenge nation, where all Kenyans want to challenge every system, all Kenyans want to build every system, all Kenyans want to be part of every system.

7:54That is why Kenyans can sit down and decide, this is what we need. I'll give an example of the mobile money transfer, which was invented here in Kenya, first used here in Kenya, and most people have never trusted Kenyans because of all that. you go to Uganda, the M-Pesa thing, you know, and that was a student, not even a graduate, just a student doing such a thing. So Kenyans are tech savvy. We love challenges as Kenyans. We have the best infrastructure in terms of the internet connectivity, you know, guys are educated. So like, yeah, that is why everyone feels so comfortable bringing their jobs here in Kenya.

8:29How did you personally get involved in data leveling? So after working as a storekeeper for quite some time, I later on joined Betin, which was a betting farm here in Kenya. And two months later, it's closed down. But there's a lady I was with in college when I was doing my diploma in Ayakago, who introduced me to summer. By the time we were together in Betin, she left, then went to summer. Then after Betin shut down, I was like, tell me, how are you surviving today? And she sent me a lingerie flight and I went to summer school. Trained, then a year later, I was up. So that was in 2020. But prior to that, I had my own challenges as a person financially.

9:07And my kid had been diagnosed with cancer of the lymphatic system. This is COVID we are talking about. And I'm like, I don't have a job. So I'm supposed to get back to my pocket because flights were not in operation globally. So I couldn't afford. But then I went to talk to someone who borrowed me 1.7 million, Kenya, actually. It was equivalent of around$1 ,700. So that is what I use for medication. So when someone came in, it wasn't offering a good pay, to be honest. It was around$240. That is on cross. $240 for month. Per month. Okay. Yeah. So I had to pay the job because I didn't have an option.

9:47I had stayed at the hospital for four months. So it was me at the hospital and me at summer at the same time until my child is actually discharged. Then now this is the burden I have. Then now I need to figure out how to find solutions to the financial crisis at my place because I need to pay house rent. I need to provide for them. Basically, I need to take care of the seed child and so on and so forth. That is how I found myself at Summer. What types of pressures are you under when you're doing data labeling? like i assume you have to be really highly accurate to be very fast as well like are you paid per task or are you paid per month per hour so at summer um we were paid per month and that is why i talked of the 240 us dollars now if you're given a target for example you're supposed to calculate and know like how many tasks do i need to submit in an hour to get the target then there was what you call the occupancy, you know, how busy you are with the tool.

10:50Because most of the tools, it was a meta tool called Workplace. And at Workplace, it could, you know, deactivate your account or put you on an available mode if you go eight minutes without touching the mouse. So, like, basically, you're supposed to be cleaned on your machine for, you know, the working period. So we had KPIs. And meeting billables also, the billable hours that, you know, the company had, you know, signed with meta was, you know, a must. So you had to work within a certain time frame. KPIs were set and you had to meet the KPIs. So what types of work were you doing at Sama? Sama was basically labeling or annotation, as we call it, and mostly images and videos.

11:34Yeah, you're given a video and they're told to describe a video, for example, you're given an image and you're told to, you know, work on it in accordance to, you know, specific guidelines. For example, if you're given, let's say, pictures of people whose, you know, the picture only contains faces and they're told to identify faces, you know, you're supposed to draw bounding boxes around their faces and label that. So that is basically what I used to do at Summer. After I joined Summer, that is when now through friends and colleagues are like, guys are talking about some other companies, some other opportunities online and so on and so forth.

12:08So the only advantage we had was we transitioned to work from home. But it was during that phase that I also came across another gig that was not a good one for that matter, where we were expected to undertake pornography. And pornography in this case was like you're presented with a video and you're supposed to put yourself in the minds of the 8 billion people on earth. and you like to put tags on every frame and by frame i mean every second of that video so you have eight billion people in mind i may have somebody searching for this pornography in cuba these are the tags they can use like if you're searching doggy you know such kind of thing so you have to have all that in mind every time you're watching a video you put tags like 12 to 15 tags on every frame so you're supposed to work on pornography for eight hours a day.

13:08I did the project for eight months. So you understand that it's something I never want to talk about sometimes because it wasn't an easy thing. Watching pornography for eight hours a day and for eight months. I went for therapy for six months. It must be very sort of destabilizing to do that. So much because you get to a point where your body can't function. you were here even if someone's still naked you don't even feel it and you're here you have a wife yeah who expects a lot from you being a young family you know a young woman she also has blood well flowing you know in her veins and she expects a lot from you intimately but then you can't like do that yeah it fractured a lot of things that time and that is how most of the things we are lost in the process were you specifically labeling like what was happening in the video and and for do you know what the purpose was was it for searching or was it for like this pornography is actually illegal so we can't include it or or what i think for them it was basically for to improve viewership on their websites for example because if you enter the internet and you want to search for certain pornographic video i know there are tags you you're likely to use to you know to be able to access that, you know, kind of video.

14:33So basically it was to improve their viewership from my own assessment. Because why would I do 12 to 15 tags on one frame? It's to ensure that even if someone was thinking outside the box, they would still come back to, you know, accessing this video. Even if someone was thinking so poorly, like, they would still come back to this very, very video. That was the main purpose of, you know, having all these tags on frames. Do you know what websites these were ultimately for, or you didn't have that level of information? Most of, you know, the communication was done on a no reply mail. So you receive a no reply mail, but it has a link.

15:04Then it, you know, moves you to a certain site. And what I noticed, they had like three working sites. So today you're working on this one, the next one you're working on that, so you couldn't tell who specifically was responsible or rather who specifically was giving this work. Then once the work is done, like your account is suspended, you do not have access. So in your report, you talk about how you feel like you were training like AI companions at some point. Can you talk to me about how you got into that? Was that the same job or a different one? That was a different one. Because this now was basically chat moderation and not data labeling or data notation, let me say so.

15:48And when you talk about trying to train these chat boards, you sit down and ask yourself, how do these companies resolve disputes? because part of the things you're told from the beginning is like you're not supposed to share your personal information for example so here you are you share your personal information the system or rather it's flagged then now that is how they narrow down to you and suspend your account so that means because again you still have your messages there's a message count on you know in your dashboard and that is what they used to pay you how do you identify the message count without having the messages how do you prove the message count how do you solve conflicts So it means these messages are being stoned for future use.

16:31Right. And we are like, we are here. And these messages, there are certain strict guidelines you're given that you must have these number of characters before a message can be sent. So that means you are not allowed to send short, very short messages. You're supposed to be certain, you know, a threshold of characters on every text message you're sending. And that is why they require someone with a high typing speed of at least 43 to 50, you know, words per minute. and that is where you're like now the speed they need is to ensure that they get the very many characters they need within the one minute that is supposed to submit that text message so the response has to be within one or two minutes depending on that site you're actually on that is where they need a better thing speed yeah do you think that you were talking to real people to be honest yes because you could feel the human aspect of their conversation and uh there are times when you know when someone's looking for love you can tell if this is a real human being or not and uh most of the people who on those sides were lonely people and i really doubt if in kenya we can have people who are lonely to their you know to that extent of you know paying for such services and uh to me i would say um i think we are raised totally different as a people and generally as africans and there are certain things you rarely find in africa for example you You know, people paying subscription fee to have such services offered to them.

17:56Like someone feels lonely about whether they can pay to get company. I haven't seen it anywhere in Africa. You know, it's there. But I think they target the vulnerable and mostly the aged people. Let me say the people who have, you know, gone through a lot. Like someone suffered in an accident, for example, lost their partners. And what have you? They have had mental issues maybe and they need support. They need someone to talk to. you know, those guys were looking for love. And now all these things are all integrated on one side. And that is what you're supposed to be responding to. So a target group in Latin Africa, to be honest.

18:36Yeah. It must be US and Europe. Yeah. So you log into this job where you're doing a chat, you're chatting with people. I mean, what sorts of things are you like asked to do? This job requires a lot of creativity. and fast thinking. Because if I'm talking to a man, I'm supposed to act with a woman. If I'm talking to a woman, I act with a man. If I'm talking to a gay, I need to act like one. If I'm talking to a lesbian, I need to act like one. So it requires a lot of creativity in switching of all these conversations. Did you have to be explicit with these people? Yeah. Were you texting them, essentially?

19:16That was basically what we were doing. There are guys who come here And they need someone to search them. And the site provided even pornographic stickers. So you said that, you know, having to watch pornography for eight hours a day desensitized you. It was not a good job. This seems like almost another level of that because you're like having to participate back and forth. I mean, was this difficult for you? The pornographic one was a different one. Right. It was a different one. I'm saying, but you did that and then you did the sexting. I was doing these at the same time. You were doing them at the same time.

19:52I was doing the same time because I have a shift between four and nine for the church moderator job. Then I'm having the other one after nine to around three or four a.m. in the morning. That is how I used to work. I'm experiencing these two jobs, having these two jobs with different experiences, but all of them are so demanding. I have to do the summer shift during the day. Basically, that is how we ended up. Some of us never sleeping for at least we used to work for at least 18 hours a day and that is where some of us are still suffering from insomnia to date and when i told you i have gone for like three days without sleep basically a minute because sometimes you're like there you switch on the lights but you can't sleep and you realize it's already 5 a.m in the morning and you can go to bed at 5 a.m so you're like supposed to pick up something else and do it so i we've most of the guys who've been They especially have a problem with sleep, most of them.

20:45I happened to go to, there is an institution called Faraja Cancer Support Center in Park Coffins, where I engaged one of the therapists who really helped me big time. And I have always said Faraja Cancer Support Center has been of immense help to me as a person, because some of the things they took me through to help me, I don't think everyone would have done that. But it was one of the best services I ever got for that year because they stood with me. They were there for me. And I think I got to access them because I was a caregiver to a child with cancer. That is how I actually got to interact with them.

21:26And as I was like, no, I need a solution to this. Yeah. So I take my child. I go there for therapy as a caregiver of a cancer patient. And at the same time, I have to explain to them this is what I'm going through at a personal level now. So that is how now I got him. Yeah. Did any place that you work for ever offer you any sort of mental health support at all? Did they even talk about it ever? No, at some point we had that. I wouldn't say it was in there, but sometimes you see you go through problems where even the therapist doesn't want to tell you. They don't even want to tell you. Not to look like I'm looking down upon them, but like, okay, the issues you go through, when you go there explaining it to them, they're also shocked.

22:11So you're like the one who's doing counseling to them, not them doing it to you. So sometimes there was that gap because now they did it from a general point of view, not from a worker point of view, because what I'm experiencing is not what they have experienced. You see, like, I'm trying to advise you on how to curb drugs, for example, and I have never used one. How the effect of, you know, cigarettes, you know, affect someone and they have never smoked. I can't tell you how it tastes unless I taste it. so for us I felt like those guys needed to have a space because they used to do it from a general point of view but if they label data for example and not just labeling data but a graphic one for that matter they would understand what we talk about so I would personally I would have suggested they go through you know data labeling first as part of equipping them with the knowledge of what is in the pipelines or what is happening to these people for them to be able to advise people accordingly.

23:09Do you regret doing this job? Or do you feel like it's something that you just needed to do in order to, you know, provide for your family and your child? I didn't have an option. If I did, I couldn't have done that, but I didn't have an option. I really had a burden of, you know, a financial burden that I needed to sort at that time. So the best I could have done is just accept the job. So I think that is the vulnerability that most people are going through. and that is what that is why most of these companies are taking advantage of you know some people because under normal circumstances I don't think you take that job today I wouldn't take it I wouldn't take it whatsoever because I know the damages that this job can actually bring to any human and also that is the reason why we formed the Data Labelers Association we need to address some of these problems and I will remind them that we're here and they will feel our impact.

24:10Unless they do things the right way, I'm going after them. And the good thing is I'm doing it from a point of experience, not assumption. I have been through this, I know what I'm talking about and I know how to approach it. So I know today they might try to ignore who we are, but at one point they'll call us on these tables, we'll sit down with them because we will ensure that we have the right protections in place, the right policies in place and we must be part of the policymaking process. We are not going to be left behind because we understand this space better than any other person. Because I went through that mess.

24:46I understand what it is and understand what can be done to find a solution to this problem. Yeah.

24:57Incogni is one of those tools that I really wish I had years ago. I mean, you've been following our coverage. Data brokers, hackers, harassers, and bad actors form a terrible ecosystem that puts you and your loved ones at risk. Scams and identity theft don't start with some dramatic hack. They start with your personal data being exposed everywhere online. Your home address, your phone number, your email, your ethnicity, your job, relatives' names, the list goes on and on and on. All this information gets collected, repackaged, and sold across people search sites, shady directories, data broker databases, and most importantly, all this can just randomly show up on the general internet.

25:34And that exposure is exactly what scammers use to target you. That's where Incogni comes in. It's hard to disappear on the internet these days, but Incogni can help. Incogni tracks down and removes your personal data from hundreds of data brokers automatically. And it focuses on making sure you don't need to repeat the process when your data pops back up. Its custom removals feature is also critical. If there's a specific website exposing your information, maybe it's a random directory, a search result, or something you found yourself, you can send Incogni the link. A dedicated privacy expert will then handle the takedown request for you.

26:10That's not just automated removal. That's personalized protection. And the more information disappears from the web, the harder it is for bad actors to target you in the first place. If you want to take control of your personal data, go to incogni.com slash 404media and use code 404media to get 60 % off an annual plan. That's incogni.com slash 404media, code 404media for 60 % off. Again, go to incogni.com slash 404media and use code 404media to get 60 % off and start protecting yourself today. Starting something new is terrifying, especially if you're doing it yourself. At 404 Media, we'd known each other and trusted each other for a long time.

26:51It was great to be able to rely on one another. We knew how to do journalism, but we didn't know much about the business. Thankfully, we didn't have to go it alone, and you don't either. Shopify is your built-in business partner. Shopify simplifies and streamlines so many of the things you need to run a business, but it's still an incredibly powerful platform. Shopify is the commerce platform behind millions of businesses and 10 % of all e-commerce in the US, including 404 Media's merch shop. With Shopify, you can build a beautiful store using ready-to-use templates so you can have your store set up in minutes.

27:24It helps you list, manage, and fulfill orders, create email and social campaigns like you've got a full marketing team. Well, basically, it's your all-in-one business platform that will help make sure 2026 is the year you actually start and run a successful business. All of this doesn't really matter if your store doesn't convert though. Thankfully, Shopify is all tied together with that iconic purple ShopPay button. It's the best converting checkout on the planet. Start your business today with the industry's best business partner, Shopify, and start hearing today. Sign up for your$1 per month trial today at shopify.com slash media.