Daniel sent us this one — he's been watching the AI industry from the inside, and he's noticed something ugly. While tech companies announce mass layoffs with one hand, they're quietly hiring an invisible workforce with the other. Hundreds of thousands of people, mostly in Kenya and India and the Philippines, doing the repetitive grunt work that makes AI look intelligent. Labeling images, ranking chatbot responses, filtering toxic content. Daniel wants to know where this is actually happening, what kind of annotation we're talking about, and the big uncomfortable question — why does this work exist in the first place?
The contrast is genuinely jarring. You've got Sam Altman on stage talking about artificial general intelligence, and somewhere in Nairobi a twenty-three-year-old is getting paid a dollar sixty an hour to teach GPT not to say racial slurs. Same pipeline, different universe.
The glamour and the grind. One's got a TED talk, the other's got repetitive stress injuries.
That's the thing. AI is sold as magic — emergent properties, sparks of general intelligence, models that supposedly teach themselves. But peel back the curtain and you find what Mary Gray and Siddharth Suri called "ghost work." Millions of people doing tasks so mundane and repetitive that companies would rather pretend they don't exist.
Daniel's question gets at something specific here. Mass layoffs are happening across tech. Meta, Google, Microsoft — tens of thousands of people shown the door. But in the same breath, the annotation industry is exploding. The global market was valued at two point two billion dollars in twenty twenty-three, projected to hit thirteen point seven billion by twenty thirty. That's not a coincidence.
No, it's a structural shift. Companies are replacing expensive domestic employees with cheap offshore labor, but they're calling it "AI training" instead of outsourcing. The work hasn't disappeared — it's just been reclassified, moved overseas, and stripped of benefits and job security.
You get this strange two-track system. Silicon Valley engineers making three hundred thousand a year building the architecture, and a parallel workforce making two dollars an hour teaching the model what a stop sign looks like. Same product, completely different relationship to it.
Daniel, being in the industry himself, sees the tension. He called it a disappointing side of AI, which is putting it mildly. This is the foundation the whole edifice sits on. Without annotation, most of these models don't function. But the people doing the work are treated as disposable inputs rather than workers.
Which brings us to the first thing we need to unpack. What is data annotation actually, and why is it suddenly this booming global industry? Daniel specifically asked whether we're talking about supervised learning labeling or reinforcement learning from human feedback. The answer is both, and they're different beasts with different implications.
So let's start with the basics. Data annotation, at its core, is humans teaching machines what things are. Want a self-driving car to recognize pedestrians? You need humans to draw bounding boxes around pedestrians in millions of images. Want a speech recognition system to understand accents? You need humans to transcribe thousands of hours of audio. Want a content moderation system to flag hate speech? You need humans to read the hate speech and label it.
The "why" is pretty straightforward. These models are pattern-matching machines. They don't understand anything — they learn statistical correlations between inputs and outputs. To learn that pattern, they need examples where someone has already said "this input maps to that output." That's the ground truth. Without it, the model is just staring at noise.
Despite all the advances in self-supervised learning, most production systems still need human-labeled data for fine-tuning and safety. The self-supervised part gets you a base model that can generate plausible text. The human-labeled part is what makes it useful and safe. So the annotation isn't a temporary crutch. It's baked into how these systems are built.
That's the first misconception worth busting. People assume AI models learn on their own, that they're somehow ingesting the internet and emerging with understanding. The reality is that someone — many someones — had to manually label enormous datasets. ImageNet, which kicked off the deep learning revolution in twenty twelve, required humans to label over fourteen million images. That's not automated. That's years of human labor.
Daniel's first question — why is this being done — has a technical answer that's almost boring in its simplicity. Machines can't learn from raw data the way humans do, so we pay other humans to structure the data for them. The interesting part isn't the technical necessity. It's who does the work, under what conditions, and why the industry works so hard to keep them invisible.
Let's dig into that. The companies building these models — OpenAI, Google,, Anthropic — they don't typically hire annotators directly. They contract through intermediaries like Sama, iMerit, Appen, Scale AI. These middlemen set up operations in countries with low labor costs and weak labor protections. Kenya, India, the Philippines, Uganda, Nepal.
The wages are shockingly low. Rest of World did an investigation finding Kenyan workers labeling data for Chinese AI firms at one to two dollars per hour. These are people with university degrees, in some cases, doing cognitively demanding work for less than the price of a cup of coffee in San Francisco.
One to two dollars an hour. These workers are the human foundation of systems worth billions of dollars, and they're making less than minimum wage in the countries where the AI companies are headquartered.
The Time magazine investigation from twenty twenty-three is the one that really broke this open. OpenAI was using Kenyan workers through Sama to label toxic content for ChatGPT's safety filters. The workers were exposed to graphic descriptions of violence, child abuse, sexual assault — the worst corners of human expression — for take-home pay of about a dollar thirty-two to two dollars per hour. These were the people making ChatGPT safe enough for public release.
The safety filter that makes ChatGPT polite and inoffensive was built by people being traumatized for poverty wages. That's not an edge case. That's the production pipeline.
It's worth naming the mechanism here, because Daniel asked about the distinction between supervised learning annotation and RLHF. The toxic content labeling is mostly supervised learning — you need humans to say "this is hate speech, this is violent content, this is child exploitation material." But RLHF, reinforcement learning from human feedback, is a different beast.
Explain RLHF for people who haven't stared at the acronym.
With RLHF, you're not just labeling data — you're training the model's behavior. You show a human rater two or more outputs from the model and ask them to rank which one is better. More helpful, more honest, less harmful. Those rankings create a reward signal. The model learns to produce outputs that humans prefer.
Instead of "this image contains a pedestrian," it's "this response is more helpful than that response.
And this is the technique that made ChatGPT feel conversational and cooperative rather than just a text-completion engine. The thing is, RLHF is more subjective than bounding boxes. A stop sign is a stop sign. But "which response is more helpful" depends on cultural context, personal judgment, and the specific instructions given to raters. Which means the biases of the raters, and the biases of the instructions they're given, get baked directly into the model's behavior.
The raters are often working under intense time pressure, with minimal training, for very low pay. A Forbes India piece describes India's "annotation army" — over a million workers, many in rural areas, doing piecework for US and Chinese AI companies. These aren't AI researchers carefully evaluating model outputs. These are people trying to hit quotas so they can make rent.
The scale is staggering when you add it up. Kenya, India, the Philippines, Uganda, Nepal, Venezuela — anywhere with an English-speaking or multilingual workforce and low wages. The annotation industry has become a kind of shadow tech sector, employing millions globally but with none of the prestige or compensation of Silicon Valley.
Daniel's framing is right — this is being portrayed in media as a new wave of miserable work. The "AI serf" narrative. But I think there's a danger in that framing too, because it can slide into fatalism. The specifics matter. The wage data, the contracting structures, the psychological toll — those are what make it possible to demand something better.
Let's talk about the psychological toll, because it's one of the most underreported aspects. Workers doing RLHF and content moderation are exposed to traumatic material. The Conversation published an investigation documenting PTSD rates among Kenyan content moderators. These are people who spend eight hours a day reading descriptions of torture, watching beheading videos, reviewing child exploitation material — all so that you and I don't have to see it when we use ChatGPT.
The mental health support, where it exists at all, is often laughably inadequate. A counseling session once a month. A wellness webinar. Meanwhile the worker is going home every night with images they can't unsee.
Sama eventually canceled its content moderation contract for OpenAI, citing the psychological toll on workers. Which sounds responsible until you ask: who picked up that contract? Some other intermediary, with some other workforce, probably in a country with even fewer protections. The work doesn't disappear. It just moves somewhere with less scrutiny.
Let me pull on a thread here. You mentioned that thirty percent of ImageNet labels contain errors. That's from a twenty twenty-four study. What does that actually mean for the models trained on that data?
It means the models inherit those errors. Mislabeled medical images lead to diagnostic errors. Mislabeled pedestrian detection data leads to autonomous vehicle failures. Mislabeled hate speech data leads to content moderation systems that over-censor marginalized groups or under-censor genuine threats.
The quality of annotation directly determines the reliability of the AI system. And we're paying people a dollar sixty an hour and wondering why the labels aren't perfect.
There's a fundamental tension here that the industry doesn't want to acknowledge. High-quality annotation requires well-trained, well-compensated workers who have the time and mental bandwidth to make careful judgments. But the economic model pushes toward the opposite — minimally trained, maximally exploited workers racing against quotas.
It's the classic outsourcing contradiction. You want it cheap and you want it good, and the market optimizes for cheap.
There's a geopolitical dimension that makes this even more complicated. China is building its own annotation infrastructure in Africa and Southeast Asia, competing directly with US firms for the same labor pools. That's not just an economic story. It's an AI arms race playing out through labor markets. Two superpowers bidding for the cheapest possible human intelligence, and the workers are caught in the middle making less than either country's minimum wage.
Then there's the automation paradox, which is almost darkly comic. The workers training AI systems are, in many cases, training the systems that will eventually replace them. Companies are actively researching ways to automate annotation — using synthetic data, using active learning where the model identifies which examples it's most uncertain about, using larger models to label data for smaller ones.
The MIT Technology Review reported in April of this year on humanoid robots being trained via gig workers. Human annotators in the gig economy teaching robots how to move and manipulate objects, and those robots will eventually replace warehouse workers and factory workers. The annotators are literally training their own replacements for poverty wages.
It's an ouroboros of exploitation.
That's the phrase.
Let me try to pull together what we've laid out so far, because Daniel's prompt has several layers and I want to make sure we're answering him directly. He asked where this work is happening — predominantly Kenya, India, the Philippines, with significant operations in Uganda, Nepal, Venezuela, and elsewhere. He asked whether we're talking about supervised learning labeling or RLHF — the answer is both, often happening in the same facilities, sometimes by the same workers. And he asked why it's being done — because AI models fundamentally require human-labeled data for fine-tuning and safety, and the economics of the industry push that labor toward the cheapest possible workers in the least regulated markets.
I'd add that the "why" has a second layer. It's not just technical necessity. It's also a choice about how to structure the industry. Companies could invest in better automated labeling techniques. They could hire domestic workers at living wages. They could build annotation cooperatives where workers have ownership stakes. They don't, because the current system is profitable and the workers are invisible to consumers.
The invisibility is the point. If ChatGPT users knew that the model's pleasant demeanor was built by a twenty-six-year-old in Nairobi being paid two dollars an hour to read descriptions of child abuse, they might feel differently about the product. So the industry keeps the annotation workforce hidden behind layers of subcontracting and nondisclosure agreements.
OpenAI isn't hiring annotators. OpenAI is contracting with Sama, who subcontracts to a local firm, who hires workers on short-term contracts. By the time you're four layers deep, nobody knows who's ultimately responsible for working conditions.
Daniel, working in AI himself, sees this from the inside. He called it disappointing. I think that's the professional's perspective — you know how the sausage is made, and you wish it were different, but you also understand the technical constraints that make annotation necessary in the first place.
That's the tension we should sit with. It's easy to condemn the whole enterprise from the outside. It's harder to figure out what a better version looks like when you understand why the current version exists.
Which is where we should go next. The economics, the human cost, and what — if anything — can be done about it.
Why don't companies just automate this? If AI is so advanced, why can't AI label its own training data?
The snake eating its own tail. Tempting, but it turns out the tail doesn't taste very good.
Automated labeling exists — it's called weak supervision or programmatic labeling — but it has a fundamental limit. You can use a model to generate labels, but you need humans to verify those labels are correct. Otherwise the errors compound. A model trained on noisy automated labels produces noisier outputs, which get used to label more data, and you spiral into what researchers call "model collapse.
The human is the ground truth check. You can't bootstrap your way past that.
And for high-stakes domains — medical imaging, autonomous driving, content safety — the cost of a wrong label is catastrophic. You don't want a self-driving car trained on data where "pedestrian" was auto-labeled by a model that's ninety-five percent accurate. That five percent error rate means people die.
Which brings us back to the offshoring question. If you can't automate the humans away, you find the cheapest humans you can.
The tradeoffs here are brutal. Companies want cheap, fast, and accurate. You can pick two. The market optimizes for cheap and fast, which means accuracy suffers. That thirty percent ImageNet error rate is what happens when you pay people by the image and expect them to maintain perfect attention for eight hours straight.
You get models trained on sloppy data, making sloppy decisions, and the worker who made the labeling error gets blamed — or more likely, just doesn't get their contract renewed. While the company that set the impossible quota faces zero consequences. The subcontracting chain is designed to diffuse responsibility.
The story doesn't end with economics. The human cost of this invisible workforce raises deeper questions about the kind of AI we're building.
The annotation boom is creating something we don't have a good name for yet. A global underclass of workers who train the systems that will eventually replace them. It's the gig economy with even less visibility — Uber drivers know they exist in the public imagination. Nobody pictures the person who labeled fourteen thousand images of crosswalks.
There's a phrase that keeps coming to mind. Not just burnout or stress — the damage that comes from being forced to participate in something that violates your sense of right and wrong. These workers know what they're building is valuable. They also know they're being exploited to build it.
It gets weirder. Companies are pouring research money into automating annotation itself. Synthetic data, active learning, using large models to label data for smaller ones. The goal is to eliminate the human annotator entirely. So the workers are in a race against the very automation they're enabling. Train the model well enough and you work yourself out of a job. Train it poorly and you get fired anyway.
Neither outcome comes with a severance package.
The question becomes what you do with this knowledge. Daniel works in AI. He's seeing this from the inside. And I think the worst thing would be to just shrug and say "that's how the industry works.
The shrug is the enemy here. The whole system depends on nobody looking too closely. If you're an AI practitioner — Daniel's position — there are concrete questions you can ask your vendors. What are your annotators paid? What's the hourly quota? Is there mental health support, and is it actually accessible or just a checkbox?
You have to ask those questions directly, because the subcontracting chain is designed to make the answers hard to find. "We contract with a firm that contracts with a firm" is not an acceptable answer. If you're buying labeled data or RLHF services, the labor conditions are part of the product, whether you acknowledge them or not.
The conflict mineral analogy keeps coming up. We decided as a society that companies should have to disclose whether their supply chains involve forced labor or funding armed groups. We haven't made that leap for AI supply chains yet, but the logic is identical. The raw material here is human attention and judgment.
There's a policy dimension. Algorithmic accountability laws are starting to emerge — the EU's AI Act has transparency requirements — but they mostly focus on model behavior, not labor conditions. Requiring companies to disclose their annotation supply chains the way they disclose conflict minerals would change the economics overnight. If OpenAI had to publicly list every subcontractor and wage rate, the two-dollars-an-hour thing becomes a brand problem, not just a Time magazine investigation.
For listeners who aren't building AI systems or writing legislation, there are organizations doing the work of documenting these conditions and pushing for standards. The Partnership on AI has a working group on fair labor practices in data enrichment. There's an AI Labor Collective organizing annotators across multiple countries. Supporting those efforts is one of the few levers an individual has.
I think there's also a more uncomfortable question for the rest of us. Every time you use ChatGPT or Claude or whatever comes next, you're benefiting from this labor. The polite refusal, the safety filter, the helpful tone — those weren't coded by an engineer in San Francisco. They were trained by someone making two dollars an hour in Nairobi.
That doesn't mean you stop using AI tools. It means you stop pretending the magic is free. Someone paid for it, and it wasn't the company.
Where does this actually go? We've got millions of people doing this work right now. The market's projected to sextuple in seven years. But the same companies funding that growth are also funding research to automate it away. What happens when they succeed?
Two futures, neither of them comfortable. In one, automation does eliminate most annotation jobs, and you've got millions of workers — many in economies where this was the best option available — suddenly without income and without a safety net. No severance, no retraining, no transition plan. Just a notice that the contract isn't renewing.
The invisible workforce disappears, and nobody notices because they were never visible in the first place.
In the other future, annotation doesn't go away but shifts. As models get better at basic labeling, the remaining work becomes harder and more specialized. You need domain experts — doctors labeling medical images, lawyers evaluating legal reasoning, engineers assessing code. Fewer workers, better paid, but the barrier to entry goes way up.
Which sounds better until you ask what happens to the million people in India's annotation army who don't have medical degrees. They don't graduate into the specialized tier. They just get left behind.
There's a third possibility that almost nobody in the industry is talking about seriously, but I think it's worth naming. Worker-owned annotation cooperatives. Instead of the four-layer subcontracting chain where each layer takes a cut, you have annotators who collectively own the operation, set their own wages, and share in the profits.
Like a guild model. Skilled laborers who control their own conditions.
It sounds utopian, but there are small experiments happening. Some annotator collectives in Kenya and India are trying to cut out the intermediaries and contract directly with AI companies. The challenge is that the companies prefer the current model — it's cheaper and it diffuses responsibility. But if enough practitioners demanded it, the economics could shift.
The practitioners are the leverage point. Daniel and people like him. If AI engineers started treating annotation labor as part of their professional responsibility rather than someone else's problem, the whole subcontracting edifice gets a lot harder to maintain.
That's really the thing I'd leave people with. If you're building AI, the humans behind the data are not an externality. They're part of your supply chain. You don't get to pretend the model trained itself.
The magic has a price tag. Someone's paying it, and it's not the company.
Now: Hilbert's daily fun fact.
Hilbert: In the late Victorian period, a Russian expedition to Kamchatka discovered that the indigenous Itelmen people had independently developed a logographic writing system for recording trade with passing ships — a script that, like Linear B, began as pictograms before evolving into phonetic symbols, but unlike any known writing system, was invented in direct imitation of the markings left by bears clawing territorial boundaries into birch trees.
Bears invented writing.
Thanks to Hilbert Flumingtop for producing. This has been My Weird Prompts. If you want to send us a question — or yell at us about bear linguistics — email the show at show at my weird prompts dot com.
Remember the humans behind the data. See you next time.
