# ‘I pretend to be a cannibal and watch porn for 8 hours a day – it’s all part of my job’ > Source: > Published: 2026-08-09 12:00:00+00:00 Joan Kinyua, 38, has thought a lot about how she’d kill, skin, cook and eat a human, all while her two-year-old girl played in the next room. But for Joan, who lives in Nairobi, [Kenya](https://metro.co.uk/tag/kenya/), this was once all part of her job. She’s a former data labeller, the humans behind [AI](https://metro.co.uk/tag/artificial-intelligence/). ‘People are really strange,’ Joan, the co-founder of the [Data Labelers Association (DLA)](https://datalabelers.org/), which has nearly 900 members worldwide fighting for better pay and perks, tells **Metro**. ‘These are things that I thought nobody ever thought about, especially publicly asking on laptops. ‘When I learnt this was all helping tech bros to make sure people don’t use AI to harm others… I couldn’t touch a laptop or [social media](https://metro.co.uk/tag/social-media/) again.’ ## What is a data labeller? AI doesn’t come ready-made [knowing the difference between a tree and a car](https://metro.co.uk/2026/04/11/ai-will-lie-cheat-disobey-humans-protect-kind-study-shows-27875177/), or why you shouldn’t tell someone how to cook a human. That’s where data labellers come in. Also called data annotators or AI tutors, they ‘tag’ text, images and videos to teach algorithms. They, for example, annotate photos of roads – that smudge is a child, that red circle a traffic light – so [robotic cars](https://metro.co.uk/2025/10/15/waymo-driverless-taxis-coming-london-safe-24433268/) can drive. Or, as Joan did, label gruesome surgical videos so [AI can diagnose health issues](https://metro.co.uk/2026/01/08/chatgpt-wants-go-health-advice-hand-medical-records-26150564/). The same work [keeps AI safe to use](https://time.com/6247678/openai-chatgpt-kenya-workers/). ‘A cannibal cannot ask what human flesh tastes like, or prompt how to skin a person alive,’ Joan says. ‘Now you can’t ask stuff like that. It’s not because they taught themselves, but somebody somewhere sat down and thought of the weirdest, most gruesome, most horrific things a person could ask.’ Rather than cannibalism, former data worker Michael Geoffrey Asia [tagged pornography for nearly a year](https://metro.co.uk/2025/07/25/checked-porn-sites-work-one-important-reason-23739712/) so it’s searchable: What do the actors look like? What position are they in? Among other things. This left the secretary general of the DLA with insomnia, PTSD and trouble having sex. ‘Having to witness three huge guys molest a 13-year-old, that remained with me,’ Michael says. ‘I have a daughter. It put me in a position where I couldn’t trust any other man close to my child. A fear that you cannot fight, even from within.’ Some data labellers don’t know their employers as they’re posted to job boards via a ‘middle man’, allowing companies to keep work private from snooping competitors. Michael even signed non-disclosure agreements, preventing him from discussing his work. With few ways to get support, Michael finally received therapy through a medical centre as he was caring for a [cancer](https://metro.co.uk/tag/cancer/) patient. The author of [The Emotional Labor Behind AI Intimacy](https://data-workers.org/michael/?ref=404media.co)’s next gig was ‘training’ sexbots. ‘There is a man on the [other side trying to seduce me](https://metro.co.uk/2025/03/04/elon-musks-chatbot-phone-sex-users-will-hit-global-birth-rates-22659874/), believing I’m a woman,’ he says. ‘Or I’m supposed to act gay, or work with lesbians. ‘The morning after, you go to church. You’re praying, asking, “Did I do the right thing? Am I pretending?”‘ Joan also ‘trained’ companion bots, which create [lifelike friends and partners](https://metro.co.uk/2026/05/16/i-fell-love-ai-chatbot-saved-real-life-marriage-28165749/), but she believes she was chatting with real people. ‘Somebody was 100% in love with it,’ Joan says. ‘I viewed the human as vulnerable; they opened up about how they’re struggling to find love, telling you their deepest secrets.’ ## How do they get jobs? **Metro **was shown a job board filled with dozens of data labelling positions. ‘Inspector Genital 🕵️,’ read one listing. ‘In this job, you will select Real Genitals, Animated Genitals or No, depending on if the image contains any genitals.’ Do that 1,000 times and you’ll earn $0.30 Other gigs included: - Naming a film in a short video clip 1,000 times for $0.15. - Determining whether a prompt and the AI-generated image are ‘controversial’ 1,000 times for $1.50. - And ‘Where Have You PEEN?’, which similarly asked the user to identify male genitalia. Many asked the user to complete a task within a minute, **Metro **saw. ‘Even if you’re given, “how do you molest a two-year-old,” despite having a two-year-old in the house, the task is timed,’ Joan says. ## ‘AI labellers are refused breaks’ The [Data Workers’ Enquiry](https://data-workers.org/), representing workers from nine countries, has [found ](https://data-workers.org/wp-content/uploads/2024/09/Fasicas-Report-Accessible-1.pdf)that AI tutoring can increase mental health issues and drug dependency. Member Camilla Salim Wagner isn’t surprised by Joan and Michael’s experiences. AI is seen as [Silicon Valley magic](https://metro.co.uk/2026/07/18/quantum-computing-thing-thatll-even-bigger-smarter-ai-28540469/), meaning it’s easy to forget the humans behind it, he says. ‘If it were widely known that workers who make AI systems are routinely refused breaks during the day, forced to work uncompensated overtime and exposed to such a level of toxic or disturbing content that they carry psychological scars for years, all while [still having to worry if the next paycheck will cover their bills](https://apnews.com/article/kenya-meta-facebook-sama-content-moderators-902716b8037bc4063dd1101e2dbd4c3a), this would significantly harm the reputation and credibility of AI companies,’ Wagner adds. The research intern at the Distributed AI Research Institute says tech giants look to African nations for data labelling to cut costs. The continent has scores of highly-skilled tech workers, and while many countries have strict labour laws, [not all do](https://workforceafrica.com/labour-laws-in-africa-guide-for-global-hr/). The data collection and labelling market was valued at [about £16billion ](https://www.precedenceresearch.com/data-labeling-solution-and-services-market)last year and is projected to balloon to £88billion in 2034. ‘We’re just being used to make billionaires and trillionaires,’ says Joan. Still, Michael feels that the work data labellers do is human – even if the result is for something that is anything but. ‘Everyone is so determined to train machines how to respond with empathy,’ Michael adds. ‘But no one has ever trained people to love people like us, who have been working on this information all along.’ **Get in touch with our news team by emailing us at webnews@metro.co.uk.** **For more stories like this, **[ check our news page](https://metro.co.uk/news/). MORE: [People say they hate AI slop — but do you really?](https://metro.co.uk/2026/08/09/people-say-hate-ai-slop-really-29293618/?ico=more_text_links) MORE: [I was convinced I was going to meet my soulmate – but I had been lied to in the worst way](https://metro.co.uk/2026/08/02/thought-going-meet-soulmate-lied-worst-way-29150998/?ico=more_text_links) MORE: [Google U-turns on AI Earth image feature after users ‘collapse Eiffel Tower’](https://metro.co.uk/2026/08/01/google-u-turns-ai-earth-image-feature-users-collapse-eiffel-tower-29274009/?ico=more_text_links)