In March 2018, an Uber self-driving car struck and killed Elaine Herzberg as she pushed her bicycle across a dark road in Tempe, Arizona. The car’s sensors detected her six seconds before impact, but the system could not decide what it was seeing. Unknown object. Vehicle. Bicycle. Unknown object. Human. Bicycle. The classifications flickered and shifted while Herzberg remained in the car’s path. By the time the model settled on a category, she was already dead.
The system had never been taught that humans cross streets outside designated lines. It knew crosswalks. It knew zebra stripes, painted intersections and the precise geometry of pedestrian lanes. But it did not know real life. It did not know that people take shortcuts, that they jaywalk when they’re late, that survival often means breaking the rules designed by people who don’t have to walk home in the dark.
As an AI researcher, whose work involves engaging data workers – the workers who label, annotate and moderate the data and content used to train AI models – there is a sentence I often open with in my talks, a provocation lobbed at the altar of technological hype: ‘Machines are stupid. Humans are deranged.’ People laugh nervously. The kind of laugh that betrays recognition. And the more I research AI, the more I sit with the testimonies of the people who make it function, the more this line feels less like a joke and more like a diagnosis of our digital present. Because what we call ‘artificial intelligence’ is neither artificial nor intelligent. It is human labour and human cruelty wrapped in an interface that says ‘I’m smart’.
Humans are not stupid. Our cognition is messy but adaptive. A child who has never seen a jaywalker will still understand that a person can cross a road anywhere. Machines cannot make this leap. They cannot improvise. They can only perform the behaviours they have been explicitly taught, pixel by pixel, label by label, by someone, somewhere, clicking and categorizing for cents per task. And ‘somewhere’ follows a familiar map: data workers in Nairobi, Caracas, Manila, Damascus, São Paulo.
We call this ‘machine learning’, but the term itself is propaganda. ‘Learning’ suggests autonomy, curiosity, understanding, a mind grasping something new. But when we say a machine ‘learned’ to recognize a cat, what actually happened is this: a worker in Lebanon clicked on 10,000 images, marking which ones contained cats. Another worker in Kenya wrote text descriptions: ‘orange cat sleeping’, ‘black cat on fence’, ‘white cat playing’. The machine did not learn what a cat is. It learned that certain pixel arrangements correlate with the label ‘cat’ because humans taught it. The machine will fail the moment it encounters something its trainers never labelled.
And yet we continue to use the language of intelligence and autonomy. We say: ‘The model learned’ while the actual humans doing the teaching remain invisible, unnamed, underpaid. This is not accidental language. It lets companies claim revolutionary breakthroughs while hiding an exploitative labour model. It lets us marvel at artificial intelligence while ignoring the very real, very human intelligence of workers who make these systems function.
Deranged humans
Everything horrific that you fear exists online. Every slur, every violent fantasy, every recorded act of cruelty. Somewhere in the world, someone is beheading someone, someone is violently sexually abusing a child, someone is dying in a ghastly manner. And because another human then uploads it to the internet, someone must clean it – watch it, categorize it, decide whether it violates platform policies, and remove it before you see it.
This is where the content moderators enter the story: thousands of workers in Kenya, the Philippines, India, Colombia, Brazil, Morocco – many of them migrants, refugees, single parents, who spend their days staring directly into the abyss so the rest of the world can use ChatGPT.
I’m part of the Data Workers Inquiry, a global, research initiative spanning five continents that collects testimony from data workers about their labour conditions.
I have spoken to workers who reviewed beheadings between their university classes. Workers who sifted through child abuse material before taking their own children to school. Workers who moderated livestreamed killings with a salary that could not cover rent. Workers told to ‘take a walk outside and breathe’ after their shift, as if oxygen is a substitute for structured mental health care.
These people are the backbone of the so-called AI. Without them your favourite tool would be spewing hatred and slurs like a Nazi on steroids. One of our South African co-researchers, Bothlokwa Ranta, who used to work as a content moderator, once told me something that I have never forgotten: ‘We are the janitors of the internet, we sift through the sewage so that you don’t have to.’
Listening differently
Data Workers Inquiry is inspired by Marx’s 1880 Workers’ Inquiry, that original attempt to let workers speak for themselves about the conditions of their labour. We agreed that we were not going to engage in traditional research that treats workers as data sources, interview them and publish work in journals they will never read.
We engage workers as co-researchers. They write their own analyses. They theorize their own exploitation. They control their narratives, what gets shared, how it gets framed, who benefits. Many write under their own names; some choose pseudonyms. Some write in English, others in Spanish or Portuguese. What emerges are not tales of victimhood but political dispatches from people who understand the AI system more intimately than any engineer or executive.
Jane, a Zimbabwean annotator, describes juggling three platforms simultaneously while feeding her baby: ‘I’m doing the math constantly – is this three cents worth it if it takes 45 seconds, or should I switch to the five cent task that takes two minutes? My daughter is crying and I’m clicking, clicking, always clicking.’
Abel, an Ethiopian refugee from Tigray living in Nairobi, spent 2021 labeling hate speech and ethnic slurs for AI training datasets, including genocidal propaganda targeting Tigrayans. All while his family remained trapped in the conflict zone. ‘I would see the exact phrases people were using online to call for our extermination,’ he said. ‘The same words. And I’m supposed to mark them calmly, categorize them, teach the machine what genocide looks like while I don’t know if my mother is alive.’
Wambui (a pseudonym), a Kenyan content moderator, vomits in a bathroom stall after reviewing footage too horrific to describe, then returns to her desk because she needs this job, because rent is due, because there are no better options. ‘Sometimes I can’t remember if I’m reviewing content or if I’m the content being reviewed,’ she said. ‘Either way, I have to keep watching. That’s the job.’
But alongside the accounts of exploitation and trauma runs another thread: quiet, growing, strategic resistance. Workers share spreadsheets of exploitative clients across WhatsApp groups. Building informal unions in Telegram channels. Teaching each other how to maximize earnings, spot bad contracts, protect themselves from the worst content, and increasingly, how to organize collectively for something better.
Artificial silence
Resistance is an uphill battle. The entire architecture of platform labour is designed to fragment workers, obscure responsibility and prevent collective action. AI companies don’t want you to know who trains their models because the magic disappears once you see the labour. The more ghostlike the workers, the more ‘intelligent’ the AI appears.
Tasks are deliberately atomized into micro-actions – click this, label that, moderate this image – so no worker ever sees the full scope of what they’re building. Contracts are short-term, often daily, so workers cannot build tenure or collective memory. Companies hide behind layers of subcontracting: tech giants hire staffing agencies who hire local firms who hire individual workers, creating a maze of deniability where no-one is technically responsible for wages, mental health or working conditions.
No one receives proper psychological support, even moderators reviewing the worst content humans produce. No-one is meant to stay long enough to understand the system.
What the AI industry calls ‘innovation’ is simply the old colonial script rewritten in code. The pattern is identical to historical extraction, just faster and more opaque. The aim is to extract knowledge from the global majority: data, annotations, emotional labour and the psychological capacity to absorb humanity’s worst violence. Workers in Nairobi earn $1.50 per hour training systems worth billions. The profits are centralized in Silicon Valley and Beijing and the supply chain erased through NDAs, subcontracting chains and the deliberate confusion of dubbing labour as ‘crowdsourcing’ instead of what it is: work.
AI is not a break from history; it is the latest chapter in a 500-year story of treating certain populations as raw material for other people’s progress, of rendering labour invisible so that products can appear miraculous.
Refusing to break
Machines are stupid. They require endless human correction, teaching and labour to paper over their fundamental inability to understand context, nuance, ethics, humour, grief or desperation. They cannot replace humans. They can only conceal the humans doing the work.
Humans are deranged. Not just because some humans upload horrors to the internet, but because we have collectively built an economic system where the psychological cost of ‘innovation’ is systematically offloaded onto the workers least protected by law and least visible to public consciousness. We have normalized a model where trauma is an externality, where someone else’s breakdown is the price of our convenience.
But here is the part that changes everything: Humans are also organizing.
Kenyan content moderators have taken Facebook and its subcontractors to court, forcing global conversations about labour rights and psychological harm in the AI supply chain.1
Workers in Venezuela maintain shared databases of exploitative clients, warning each other away from platforms that don’t pay or that demand impossible turnaround times. Refugees in Germany are demanding recognition, not as charity cases doing digital piecework, but as essential, skilled contributors to AI systems worth trillions. Young people in Nairobi, particularly the Data Labelers Association, are teaching each other how to navigate the platform economy, how to resist its worst demands, how to organize for better conditions even when the bosses are algorithms and the workplace is everywhere and nowhere.
Before we treat AI as a neutral tool or autonomous entity, we must confront the supply chain of trauma and silence and invisible labour that sustains every interaction.
And then there are the costs we cannot see, the personal toll hidden behind the labour statistics and organizing victories. Michael Geoffrey Asia, a text chat operator from Kenya, told me something I cannot forget: ‘Remember that an AI girlfriend responding to your loneliness might just be a man in a Nairobi slum, wondering if he’ll ever feel real love again.’
Behind the curtain
Before we panic about AI ‘taking’ jobs, we should meet the people whose jobs make AI possible in the first place. Before we treat AI as a neutral tool or autonomous entity, we must confront the supply chain of trauma and silence and invisible labour that sustains every interaction. When ChatGPT sounds wise, it is because a worker taught it not to sound deranged.
AI’s workers are not peripheral to the story. They are not a footnote or an externality or a temporary phase before full automation. They are the centre. They are the foundation. Everything else, the algorithms, the hype, the billion-dollar valuations, is branding.
The first step toward justice is visibility. Seeing the workers who have been deliberately hidden. The second is listening, letting them tell us what they know, what they need, what they’re building. The third is solidarity – standing with the people who keep our machines from revealing their stupidity and our societies from revealing their derangement.
The future of AI does not belong to the companies that market intelligence. It belongs to the workers who produce it, who are organizing in the cracks of the system, who refuse to remain invisible. And if we are paying attention, really paying attention, we might finally recognize this moment not as the triumph of machines, but as a turning point for human dignity.
The workers are already turning. The question is whether we will join them.
- Ammu Kannampilly and Humphrey Malalo, ‘Kenya court finds Meta can be sued...’, Reuters, 20 September 2024, a.nin.tl/meta
