The human cost of AI training
There’s unusual comfort in picturing that artificial intelligence (AI) finds out all by itself. We like to think of it as something separated, a machine that just occurs to understand things. The truth is far more human. A recent CBS 60 Minutes section included young Kenyans, the majority of them college graduates, working for just $2 an hour. Their job was to train AI systems. For 8 hours a day, they stared at screens. Some drew boxes around microwaves and deals with. Others circled around medical irregularities. The most disturbing tasks involved labeling violent content: child abuse, animal cruelty, pornography, even video of suicide. They did this to help AI recognize what is safe and what is not. Our machines seem respectful and valuable because someone else went through hell to teach them.
In my previous post, I focused on transformer architecture, the T in GPT. This time, I’m writing about the P, as in pretrained. In my earlier piece, I described a silent librarian inside the system, assembling significance from word vectors as collaborates that clump together because of a self-attention system despite order or position. However, these labeled words and images are what the curator memorized during pretraining. The system feels smart because somebody fed it with care, frame by frame.
Back to the CBS documentary: Some of the employees are now suing. Others won psychiatric evaluations verifying what they already knew: they were not the same anymore. One man stated he could no longer enjoy intimacy with his wife. Another said it was now easier to weep than to speak. And I needed to ask myself: if I had not seen it, would I have cared? We celebrate AI for its benefit. It writes our emails, organizes our notes, and even aids with homework. But behind each reply is a chain of labor. Unnoticeable. Outsourced.
What occurred in Kenya isn’t rare. Tech companies frequently avoid direct responsibility by using third-party companies. These intermediaries keep salaries low and protect their client-brands from blame. Files showed OpenAI paid the outsourcing firm $12.50 per hour per employee, yet the Kenyans got only $2. What disturbed me most was my own response. My first thought was utilitarian. If this helps more individuals than it harms, isn’t that a great trade? However, would I feel the same if it were my sister labeling gore to pay the bills? Or my friend watching suicide video so I could use a free chatbot? The response is not direct.
Kenya needs jobs. Its government promotes itself as the Silicon Savannah, a hub for tech outsourcing. Refusing contracts might mean fewer opportunities, even if those jobs come at a high mental expense. In my previous piece, I explained the model as a determined curator. Now I see that she stands on the shoulders of exploited human beings. Labelers feed her the information. They sort the material. They carry the psychological weight. The curator never sleeps, the ones who trained her are tired, often broken.
The problem runs deeper than just money. Our systems are built to conceal the cost. Executives do not see the harm, and software engineers do not write the contracts. The users do not understand what happens behind the scenes. So, we all pretend it is progress. Perhaps ignorance is not bliss. Perhaps it is complicity. I am not calling for a boycott. I still use AI. It helps me work. We cannot claim to be ethical if we benefit from suffering that we choose not to see.
Maybe the point is not to condemn or cancel. Perhaps it is just to see more clearly. I admit, it feels strange. Understanding all this and still using the tools. There’s a contradiction I cannot ignore. Maybe this discomfort is exactly what we need. There is something to be said for the weight of understanding. When we pretend the harm does not exist, we can use these tools thoughtlessly. But when we acknowledge the hands that trained our devices, we become more deliberate. We pause before asking AI to do work we might do ourselves. We question whether this job really needs automation, or whether we are just avoiding effort. The trouble of caring is not a bug. It is a function. It is our humanity. Unfortunately, it is also our nature to forget as time passes.
But if we are going to keep using, building, and improving these tools, then at the very least, let us not forget those bearing the weight below.
