The Ghost in the Machine: On the Human Labor We Pretend Doesn’t Exist
I spent last Saturday labeling pictures of parked cars. Not because I work in AI. I was trying to understand what it feels like to be the human inside the machine. Three hours later my eyes hurt, my wrist ached, and I had classified about eight thousand images of vehicles in various states of not moving. The pay would have been twelve dollars if this had been a real gig.
Somewhere, someone is doing this for real. They are labeling the stop signs and lane markings that will help an AI drive a car someday. The press will call it an advance. The company will call it a breakthrough. Nobody will call it what it is: work, done by people, poorly paid, designed to be invisible.
The magic trick only works if you do not look at the hands.
The Hidden Workforce
Nobody has good numbers because the industry does not want good numbers. Oxford’s Internet Institute estimates there are millions of ghost workers. People who label data, moderate content, train models, test outputs. They work through platforms like Amazon Mechanical Turk, Appen, and a dozen others you have never heard of.
Pay for data labeling runs about two to five dollars an hour in the United States. In Kenya, the Philippines, and India, it can drop below a dollar. The work comes in pieces, no benefits, no job security. Workers are independent contractors, which means minimum wage laws, overtime protections, and workplace safety rules do not apply.
Content moderation is worse. Facebook, TikTok, YouTube, and Twitter subcontract to firms like Cognizant and Accenture. Moderators watch child exploitation footage, beheadings, animal abuse, and terrorist propaganda for eight-hour shifts. The pay is marginally better. The psychological toll is not. Studies show PTSD rates among content moderators that match first responders.
OpenAI used Kenyan laborers paid less than two dollars an hour to filter graphic content from ChatGPT’s training data. A Time investigation in 2023 documented workers describing nightmares, lasting trauma, anxiety that would not go away.
The Design of Invisibility
None of this is accidental.
The term “ghost work” comes from Mary Gray and Siddharth Suri’s 2019 book of the same name. They argue the invisibility is a feature, not a bug. AI companies need users to believe the technology is autonomous. The moment you see the humans behind the curtain, the magic dissolves.
Gray and Suri call this the “paradox of automation.” The more complex the system, the more human labor it requires to maintain.
The platforms double down on anonymity. Workers on Mechanical Turk are ID numbers. No collective bargaining, no grievance process, no path to employment.
This is the same logic that powers recommendation engines that became moral gatekeepers.
What Ethical Obligations Do We Have?
The answer implicates everyone who uses AI tools. If you have used ChatGPT, you benefited from this labor. The line between consumer and complicit is thin.
Better wages are the obvious starting point. OpenAI raised over ten billion dollars. Alphabet is worth nearly two trillion.
The ghost workers are still there. Still invisible. Still paid pennies to make software look intelligent. I do not have a neat answer. But the first step is being willing to see them.
The Deeper Problem
This is not just a labor problem. It is a cognitive one. When we build systems that hide the humans inside them, we train ourselves to ignore the human cost.
I wrote recently about the right to look away. This feels like the darker side of that coin.
Iris Murdoch wrote about attention as a moral act. The AI industry has built an infrastructure designed to direct our attention away from the workers who make it possible.
The same pattern appears in how online communities govern themselves to death.
The machine is not magic. There are people inside.
