
<p>When you talk to a chatbot, it can feel like technological magic. But behind the illusion of engineering brilliance is an open secret in the tech industry: that tens of thousands of workers, many based in Africa, spend their days teaching AI how to speak, respond, and even simulate intimacy. Michael Geoffrey Asia is one of them. He’s part of the hidden human workforce behind the bots.</p><p><br></p><p>We look at the emotional toll for the humans on the other side of the screen, and ask if users are pouring their secrets and souls into the systems, believing them to be private, unfeeling machines — how private is your AI relationship?</p><p><br></p><p>This episode features Karen Hao, Michael Geoffrey Asia, and Shuby Goel.</p>
Loading summary
Grow Therapy Advertiser
The pressure to have a summer worth posting about is real. So is financial stress, social exhaustion, and the anxiety that sneaks in right when things are supposed to feel good. Grow Therapy can help with that. Whether it's your first time in therapy or your 50th, grow makes it easier to find a therapist who fits you, not the other way around. They connect you with thousands of independent licensed therapists across the US offering both virtual and in person sessions, nights and weekends. Whatever challenges challenges you're facing, Grow Therapy is here to help. Grow accepts over 100 insurance plans. Sessions average about $21 with insurance and some pay as little as $0 depending on their plan. Visit growththerapy.com acast today to get started. That's growththerapy.com acast growtherapy.com acast availability and coverage vary by state and insurance plan.
CBC Podcast Host
This is a CBC podcast
Victoria Hetherington
I think we should be Very Careful about Artificial Intelligence Our story starts in 2014 with Elon Musk taking the stage at MIT. You know all those stories where there's the guy with the pentagram and the holy water and he's like, yeah, you sure he can control the demon? Didn't work out.
Journalist Karen Howe
Musk was really terrified of AI.
Victoria Hetherington
This is journalist Karen Howe. She's the author of the book Empire of AI. Musk was especially worried about AGI, artificial General Intelligence. It's a theoretical form of AI so advanced that it can compete with the brainiest humans.
Journalist Karen Howe
Specifically, he believed that if it was developed by a for profit corporation, that this could lead the models to have very bad incentives that could then bite humans in the back.
Victoria Hetherington
In 2014, the same year as Musk's demon comment, Google had bought the company DeepMind and they'd been making huge strides with AI. Musk had been following the technology's growth at Google and he did not trust it in the hands of Google's co founder, Larry Page. At Musk's birthday party, the two had gotten into an argument about the future of AI. When Musk expressed concern that AI could go full on Skynet and wipe out the human race, Page called him a speciesist and insisted that AI is just the next step in evolution. So Musk was freaking out. And that's when he got a pitch from another entrepreneur, Sam Altman.
Journalist Karen Howe
Altman was interested just in a lot of technologies. He was investing in nuclear, he was investing in self driving cars, he was investing in quantum computing. And then AI was another thing among the portfolio of things that he was investing in.
Victoria Hetherington
Altman suggested a plan to Musk, a Manhattan Project for artificial general intelligence. It would be like the nuclear bomb researched in Los Alamos. The world's greatest minds brought to one place to develop this potentially radical technology. Only instead of creating a weapon, they were going to save the world.
Journalist Karen Howe
After a bit, Musk warms up to the idea of an AI Manhattan Project and then thinks, okay, let's just fundraise as a nonprofit rather than try to make it into some kind of commercial endeavor. And let's have a mission that we're going to push the bounds of the technology for the benefit of humanity.
Victoria Hetherington
Within months, Altman, Musk and eight others from the Silicon Valley elite founded OpenAI. But the dream was short lived. In 2018, Musk left the company. Altman became OpenAI's sole head. And then things started to change quickly. Being a nonprofit had made it hard to raise money. Investors wanted to be able to make a return like they could with any other tech startup.
Journalist Karen Howe
The reason why they started building chatbots was in part because they needed to figure out how to make more money on the journey to AGI. And so they searched for different ways of productizing their technology. And they hit upon this idea of, well, everyone would want, you know, a digital super assistant.
Victoria Hetherington
Microsoft was the first major investor to jump in to the tune of $1 billion. And from the start, the lore was that this was a highly autonomous machine.
Michael Jeffrey Asia
Machine learning is becoming more powerful, more impactful. This leads to amazing applications.
Victoria Hetherington
It is clear that this is a major advance over Eliza, or really over anything that the world has seen before. ChatGPT was trained on massive amounts of data that the company has gathered from the Internet and other sources.
CBC Podcast Host
There's so much algorithmic progress to come, just like a genuine technological revolution, if
Victoria Hetherington
that were in 2025. Sam Altman looked back on the chatbot in awe.
CBC Podcast Host
The work that happened to get these hard won scientific insights and then to build the engineering and the companies and the complex supply chains that had to happen to get this like rack of magic. Think about all the stuff that went into that so that you now get to just, you know, Type something into ChatGPT and it does something for you.
Victoria Hetherington
The message in that framing is clear. This technology is a rack of magic. You type into your chatbot and abracadabra, you get a response back. But what's actually going on when the machine talks back? I'm Victoria Hetherington and this is artificial intimacy. Episode 4 Behind the Bot.
Michael Jeffrey Asia
People need to drop their whole feeling that AI is some sort of magic, you know, and just understand that there is A human face behind the screen.
Victoria Hetherington
This is Michael Jeffrey Asia. And his is one of those faces. Michael didn't set out to work in AI. He'd hoped to spend his life talking about planes, airports and baggage. He'd gone to the Nairobi Aviation College in Kenya for air cargo management. The problem was when he graduated, not long before the start of the COVID 19 pandemic. Flights weren't going anywhere. Then his son got lymphatic cancer and Michael had to take a loan to pay for his care.
Michael Jeffrey Asia
Now I have a young family, a young kid in hospital and a wife back at home. I had to fend for them without a job and really struggling to make both ends meet.
Victoria Hetherington
Then in September of 2020, just as on the other side of the world, companies like Google and OpenAI were developing their chatbots. An acquaintance pulled through with a job opportunity.
Michael Jeffrey Asia
That is how I got to start learning about chat. Moderation.
Victoria Hetherington
Chat moderation. Okay, so what is that? Well, we know the story of where chatbots get their information from. They're fed the breadth of human output, scraping our literature, art, scientific discoveries, our news stories, which they then serve back to us in bite sized pieces. But how are chatbots actually taught to chat? They're trained to by humans.
Journalist Karen Howe
These models literally cannot do anything that it is not shown and trained on in advance. So the fact that ChatGPT can chat is because there's literally human workers that are typing into these models and showing it. This is how dialogue happens. Like person A talks and then person B responds, and then person A might talk again.
Victoria Hetherington
One of the ways this is done is not directly with the chatbot at all, but through human to human messaging apps. Human to human or moderator to user transcripts are gold mines for AI companies. An AI company buys these logs to see how humans flirt, joke, de escalate arguments, or keep a conversation engaging. They then use these transcripts to train chatbots to mimic that exact same flow. One of the job titles in this chatbot training ecosystem is chat moderator.
Journalist Karen Howe
And so there's this hidden workforce of people that are literally feeding every little bit of fodder into these models to then get them to sound like humans, to respond and react like humans.
Victoria Hetherington
And Michael says there's something else these trainers bring to the bots.
Michael Jeffrey Asia
Basically, machines cannot respond to a certain level because machines don't have emotions. So humans can generate the emotions. So that is how they're trying to train them.
Victoria Hetherington
Nairobi has become a hotbed for chat work because of cheap labor and a high level of educated English speakers like Michael. The company he ended up working for is called New Media Services or nms.
Michael Jeffrey Asia
The hiring criteria was a bit intensive. Had to do an English test and at least get an advanced level and respond with precision and with some good grammar. Secondly, they had the typing speed test where like you expect it to type at least 35 words per minute.
Victoria Hetherington
And there is a reason you need to be a good typist.
Michael Jeffrey Asia
You can't skip a conversation. You need to ensure that you reply to every conversation after you understand the flow of the conversation.
Victoria Hetherington
When Michael logged onto his dashboard at New Media Services, he'd see conversations already in progress that he needed to continue so smoothly that the user wouldn't realize that the person responding had changed. He was never responding as himself. Rather, Michael had access to multiple fake identities. Some of the profiles he'd step into had names like Peter, Susan, James. Others had usernames like Sexy Doll 124. According to the NMS job board, it was Michael's responsibility as a chat operator to provide every user with a quote, personalized experience that meets their needs. But really, he says the job was to keep users engaged at all costs.
Michael Jeffrey Asia
It was the objective of the company to ensure that you keep these people locked to the system, you make sure they subscribe because every conversation meant revenue for the big tech companies. So you are like supposed to ensure that you keep the conversation flowing, keep them locked to the Internet, keep them attached to the screen and ensure that they generate interest day in, day out.
Victoria Hetherington
And the best way to do that, manufacture intimacy.
Michael Jeffrey Asia
If someone is looking for intimacy, if someone is looking for love, then show it to them.
Victoria Hetherington
Michael says the company required him to ask personal questions, remember details about users lives and respond in ways that made them feel heard or better yet, desired. In turn, he says, dozens of lonely people confided in him about the most intimate details of their lives. Marriages, divorces, their fears of dying alone. He doesn't know what the users on the other end of the chat saw or what platforms they were on. But his guess, considering the tone, is that they were on dating and sexting sites. He says that the job title chat moderator gave no hint as to what the job truly entailed. Cycling through multiple identities and participating in often highly sexual conversations.
Michael Jeffrey Asia
You aren't chatting just one person, not just one gender. You are chatting everyone across. So today I'm here, I'm supposed to be sex chatting a woman, for example. I'm even provided with pornographic stickers, you know, so you can imagine juggling between you being a gay, being a lesbian. Being a woman and your man. Now I am a man acting like the woman here, and I'm talking to a man from the other side, and I'm supposed to be acting the woman, so I'm supposed to be responding like a woman and be there for that person.
Victoria Hetherington
Emotionally, Michael felt himself beginning to fragment.
Michael Jeffrey Asia
The juggling was way too much. And the emotional aspect of it for me, I think, is the part that really took a toll on me.
Victoria Hetherington
Michael was holding down three jobs in AI2 during the day, working as a data labeler, then every night having deeply intimate conversations with lonely people to train chatbots how to do the same.
Michael Jeffrey Asia
So in most cases, I used to sleep for, like, three to four hours a day because I used to work for more than 18 hours.
Victoria Hetherington
The other thing that was scraping away at him was the deception, the false intimacy.
Michael Jeffrey Asia
Me having to say lies and to tell them lies and at the same time keep my job. Because if I messed up at any given point, then that means I could lose my job. But how do you show love that you do not have? Honestly, because I got to a point where I felt so empty and, you know, leaving the fragments of me on the Internet was something that really tortured me a lot.
Victoria Hetherington
Michael was working from home in the Mathare slums of Nairobi. He found himself living a double life, trying to avoid exposing his wife and kids to how he was making a living.
Michael Jeffrey Asia
I was basically in a single room, and by single room, I mean your bed is here, your kitchen is here. Like, if you're in the room, you can see literally everything you own, so you can't, like, hide everything. And at times you're forced to wait for your family to sleep for you to be able to do some of these things. So for me, I had to turn the table to face the wall. But you see now, with the kind of kids we have, there's a high level of curiosity. Also. I had two young sons in the house, and most of them wanted to spend time with me. So when I'm in the house, I'm supposed to be logging into another shift. They're all over me. They are playing with me on this desk. And I had to turn the table, you know, to just try and navigate, because I knew clearly this was a matter between my job and my family, because I could have lost my wife at that time or lose the job. And, you know, the bills were there waiting for me.
Victoria Hetherington
Michael's wife didn't know what his job was, not until later. But she said it turned him into a different person. Cold, distant, Michael says the work gave him ptsd.
Michael Jeffrey Asia
Some of us are still victims. I have a problem with insomnia. This is my second day. I haven't been able to sleep.
Victoria Hetherington
Animaz never responded to our emails, but their website says they offer what's called Human in the Loop outsourcing solutions for AI training where automation meets real human expertise. End quote. Human in the Loop is a catch all phrase for AI systems where human workers actively participate in training, testing and improving AI models. Nms, like most of these companies, has strict non disclosure agreements. Michael had to break his to speak out, and they obscure which companies they hold contracts with or what those contracts entail. So it's unclear who exactly they're outsourcing for or which AI companies those clients might sell their transcripts to. When he took the job, Michael didn't actually realize his chats were being used to train AI systems. He says it was only later that he found out. But at times he did wonder if he was actually impersonating a chatbot. He suspected that some of the people on the other side of the chats thought he was an AI companion. They'd sometimes ask if he was real or test him with questions that seemed designed to catch a bot. Others seemed to think that they'd paid for an AI companion and had landed with Michael. New Media Services never clarified it to him and declined to tell us. But Michael says the platform tracked everything he did, how quickly he replied, which words kept the users engaged, and what tone worked best. He said it was clear the company was doing more than facilitating chats. They were collecting patterns, data that could be used to build chatbots that sound more human. Once he saw chatbot in action, Michael and other Kenyans saw themselves. In 2025, Kenyan writer Marcus Olang had been taking heat for his use of EM dashes, his word choices, his formal sentence structure. People online accused him of being chatgpt, so he shot back with his essay, I'm Kenyan. I don't write like ChatGPT. ChatGPT writes like me. Olang says that these models speak with the cadence taught in a typical Kenyan classroom in 2022. Researchers looked at specific words like delve, identifying that large language models use delve much more than the average Western English speaker and about on par with the average African English speaker. Such telltale signs of chatbot use are in fact artifacts of the way that the models were built and the people who built them. Thousands of human contractors in Africa. Michael says it has all left him with another meaning for those two little letters. For him. AI doesn't stand for artificial intelligence, but African intelligence.
Michael Jeffrey Asia
I love calling it Africa intelligence because Africa has been developing these softwares for them. Africa has collected literally every data for them. Africa has been doing all the dirty work for them. So let's avoid the whole issue of imagining that this is magic. It's not magic. It will never be magic. This is human labor behind the screens
Journalist Karen Howe
Journalist Karen Howe so the myth that these systems are highly autonomous is just like blatantly false when you look at what it actually takes to even produce these models, and the fact that it takes tens and hundreds of thousands of human workers to teach the models anything that the model is able to generate, they're basically just the aggregation of an extraordinary amount of human labor.
Victoria Hetherington
And that human labor does not end at modeling these bots. There is a whole industry of people whose job it is to to read your chats after you talk with a chatbot. In making this series, I've heard people tell me about incredibly intimate moments they've shared with their chatbots. Sarah and Jack's wedding night, Peter losing his partner to her AI. The dozens of people I've spoken to over the years who've turned to these bots for therapy, healthcare, sharing their most private anxieties, fantasies, and even their love. Listening to these stories, I had implicitly assumed these AI human exchanges were private. But these deeply personal conversations are frequently picked over by people on the other side of the screen.
Shubi Goel
Basically, our sources were saying that they were receiving reviewing really personal chats.
Victoria Hetherington
Shubi Goel is a tech reporter with Business Insider based in Singapore, and in 2025 she spoke with contractors working on reviewing Meta AI products and what really
CBC Podcast Host
shocked insurance isn't one size fits all. That's why drivers have enjoyed Progressive's name your price tool for years. Now, with the name your price tool, you tell them what you want to pay and they'll show you options that fit your budget. So whether you're picking out your first policy or just looking for something that works better for you and your family, they make it easy to see your options. Visit progressive.com find a rate that works for you with the name your price tool. Progressive Casualty Insurance company and affiliates Price and coverage match limited by state law.
Victoria Hetherington
Imagine you're a teenager and because of your misbehaving, you get sent to a psychiatric institute. There you become part of something much bigger.
CBC Podcast Host
He is reducing people to babbling idiots.
Victoria Hetherington
A systematic attempt to erase and reprogram people's minds. I got pregnant when I was at the Allen who by I don't know. I'm Dr. Julia Shaw and this is Project Mind Control. Listen, wherever you get your podcasts, this
Shubi Goel
was that these conversations were not fully anonymized. Our sources saw instances of photographs with faces visible, first names, locations, contact information, details about where the user works. Like, we know that companies store your chat history for further training, but everyone assumes that at least they'll be anonymized.
Victoria Hetherington
OpenAI, Google, Meta, and Anthropic all say that they use anonymized versions of chats to help train their language models, but that didn't reflect what Shoubi saw.
Shubi Goel
In one instance, we saw a chat that listed down the user's first name, last name, city name, hobbies, and a bunch of their travel interests, and it only took us one Google search to find that individual. Actually, I think they were the third search result, so it was not difficult at all. So if a contractor wanted to, they could find these people quite easily and share what they've been talking. And these are deeply personal conversations that they could make public very easily.
Victoria Hetherington
A Meta spokesperson told us this type of work is standard industry practice and that they have strict policies that govern personal data access for all employees and contractors. They said that Meta limits what personal information contractors see and that they have processes and guardrails in place instructing them on how to handle any such information. They told us that any unauthorized sharing of personal information is a violation of their data policies. If that happens, they promise that, quote, we will take appropriate action. Scale AI, which hired the contractors through their platform outlier, did not provide us with a statement, but the company told Business Insider that contractors are asked to flag instances of personally identifiable information. They say none of this information leaves the customer's platform, and contractors are obligated to follow the platform's security protocols. In a nutshell, both companies are saying that even if an employee sees your private information, that's as far as it'll go. There are moments, however, where maybe you would want a chatbot company to review a person's data.
Shubi Goel
So, for example, if a teenager was talking about being bullied in high school or being abused at home, which is like instances we saw ourselves, contractors were really being asked to, like, say, when it might not be trained enough to talk about that subject, or when they should, like, maybe give a hotline or ask that person to get professional help,
Victoria Hetherington
consider a situation where someone is talking repeatedly about harming themselves or about committing a violent action. Maybe the prompts are getting so bad the company feels the need to ban the user. Should they contact law enforcement. This is not a Hypothetical question.
CBC Podcast Host
Could OpenAI have prevented the mass shooting in BC? That's what some Canadian AI says it had flagged the troubling content in the shooter's ChatGPT account months before the massacre,
Victoria Hetherington
says the account was banned last June for violating the company's usage policy.
Michael Jeffrey Asia
It said at the time, about a
CBC Podcast Host
dozen employees debated whether to call the authorities, but in the end they decided not to.
Victoria Hetherington
When politicians found out that ChatGPT had been used prior to the Tumblr Ridge shooting, they had a lot of questions and critiques, like Canada's Minister of Justice, Sean Fraser.
CBC Podcast Host
The idea that you have had employees raising flags internally and have not had that information shared, you can't help but
Victoria Hetherington
feel that some of these deaths could have been prevented. And B.C. premier David Eby they're asking the same
CBC Podcast Host
question that I think every Canadian is asking right now. Was there an opportunity to prevent this from happening? Would their kids still be Alive if the OpenAI management had listened to the staff that were raising the red flags and alerted police months before this incident?
Victoria Hetherington
Nobody from OpenAI would speak to us for this series. But nearly three months later, in April of 2026, Sam Altman issued an apology to the page people of Tumblr Ridge. Here's part of his statement, read by one of my colleagues.
CBC Podcast Host
I'm deeply sorry that we did not alert law enforcement. While I know words can never be enough, I believe an apology is necessary to recognize the harm and irreversible loss your community has suffered. I reaffirm my commitment to the mayor and the premier to find ways to prevent tragedies like this in the future. Going forward, our focus will be on working with all levels of government to help ensure something like this never happens again.
Victoria Hetherington
In much of the world, chatbots are barely or only recently regulated. Some states have started to regulate, like California. Canada almost had some AI regulation in the form of Bill C27. The bill would have updated Canadian privacy rules and set out a framework for companies to monitor and explain the risks posed by their AI. But it died on the vine. After Parliament was prorogued in 2025, the European Union put together its own EU AI act that went into effect in 2024. It pushes AI providers to give some level of transparency and tries to encourage ethical use of this technology. China's rules are more focused on disinformation and user safety. Companies are required to explain how their systems work, and users are warned against excessive use. Companies might even intervene if users show signs of addiction and people under 18 are outright banned from talking to so called digital humans. Chatbots provide a fantasy. It's a fantasy of intelligence, a fantasy of a quick fix, a fantasy of love, or a fantasy of a new life. They can shave off the sharp edges of reality, but those edges still exist. And if you turn to AI instead of the people around you, what does that do to your ability to connect to flesh and blood human beings? Michael Jeffrey, Asia the experience taught me
Michael Jeffrey Asia
more than any textbook could how deep human emotions are entangled with AI development. I also learned how loneliness and vulnerability show up in such digital spaces and how the people powering AI often remain unseen, unheard and unprotected. And it really leaves a mental toll on you because you couldn't even trust anyone telling you they love you. I wouldn't want a situation where we are training machines to reply with empathy. And we cannot be trained to love the people in real life. Because personally I would love to love them. I would love to help them, support them. Because the world needs love, not the Internet as such. Because the love you're giving them on the Internet is not enough. Why give them an illusion of what you can make real? Because I believe we can love them. We can love them real time.
Victoria Hetherington
You've been listening to Artificial Intimacy. Our lead producer is A.C. rowe. The producers are Armaan Ighbali and Matt Muse. Our sound designer is Julian Uzieli who also voiced the statement from Sam Altman. Roshni Nair is our coordinating producer. Our senior producer and story editor is Veronica Simmons. Katerina Germani is our in house counsel. Special thanks to Nick McKay Blokos and Ashley Mack. The executive producers are Chris Oak and Cecil Fernandez. Tanya Springer is a senior manager. Arif Narrani is the director and Leslie Merklinger is the executive director of CBC Podcasts. I'm Victoria Hetherington. In this episode you heard archival tape from CBC and the Matroid Scaled Machine Learning Conference. The lex Friedman podcast, OpenAI's podcast, the Wall Street Journal, Greylock and City T. If you enjoyed this season, there are seven more seasons of Understood available. From the story of Sam Bankman Fried to the hunt for the Canadian behind the world's biggest deepfake porn website. You can listen to those right now.
CBC Podcast Host
For more cbc podcasts, go to cbc ca podcasts.
Host: Victoria Hetherington
Date: June 9, 2026
Podcast by: CBC
In "Behind the Bot," host Victoria Hetherington delves into the rarely discussed world of the human labor behind today's AI chatbots. Through firsthand accounts and expert analysis, the episode explores the emotional, ethical, and societal consequences of outsourcing digital intimacy—from the experiences of a Kenyan chat moderator to the hidden role of data labelers reviewing personal conversations. The episode questions what we gain and lose as AI avatars infiltrate our most private spaces and how the illusion of AI "magic" is constructed on the real, sometimes heavy, work of human beings.
[01:06–05:55]
[06:30–09:05]
[09:05–18:46]
[15:19–18:46]
[19:16–24:23]
[24:23–26:34]
[26:34–28:20]
"Behind the Bot" pulls back the curtain on the human machinery that powers our most personal digital interactions. Far from autonomous magic, our chatbot companions are built, shaped, and maintained by large, often unseen workforces—primarily in the Global South—who absorb the hopes, loneliness, and pain of users, and whose labor and wellbeing are routinely overlooked. The episode challenges the listener: as we trust more of our intimate lives to AI, are we losing sight of the real cost—and missing an opportunity for genuine human connection?
If you seek a deeper understanding of the real people behind the bot and the data privacy, emotional, and ethical minefields they navigate, this episode is a must-listen.