The Workers Inside the Machine

EDO SEGAL: Timnit, you've done something I think is permanent: you relocated the conversation about AI from disembodied intelligence to human labor. The popular story imagines autonomous systems that learn on their own. You insist on naming what the industry keeps invisible — the data labelers, the content moderators, the annotators, many in the Global South, paid a couple of dollars an hour, sometimes pennies per task, whose work is the unacknowledged foundation of the whole enterprise. I want you to tell that story the way you'd tell it to the girl in Addis Ababa from my cold open. And Pamela — I'm going to ask you whether the historian of the founders ever saw these workers, or whether they were invisible to your method too.

GEBRU: I'll tell it plainly, because the plainness is the point. There is a content moderator — let's say in Nairobi, though it could be Manila or Hyderabad — who spends her day reading the worst material the internet produces, the violence and the abuse and the horror, so that when the girl in Addis Ababa opens a chatbot it answers her cleanly and the horror is nowhere in sight. The moderator is paid a fraction of what a worker in California earns for comparable work, often less than two dollars an hour, and she carries the psychological wounds of that exposure home with her, uncompensated, unacknowledged, frequently with no mental-health support at all. The polished, magical product rests on a hidden substrate of human suffering that the marketing never mentions. And here's the deeper point, the one that changes how you think about the whole thing: the machine doesn't replace human labor. It hides and displaces it. The annotator who taught the model is rendered invisible the instant the model performs, her judgment congealed into a product that appears to need no labor at all. We call it artificial intelligence. A great deal of it is human intelligence, captured and stripped of attribution. To see the workers is to see the system clearly.

EDO SEGAL: Pamela. Did your method see them?

McCORDUCK: No. And I want to sit in that no, because it's the most damning thing I'll say about my own life's work tonight. My method was portraiture — I wrote the field by writing the people, because I believed, correctly, that you can't understand a science without understanding the temperaments that built it. But look at whose temperaments I chose. Simon. Newell. McCarthy. Minsky. Feigenbaum. The men in the room with the funding and the vision and the names that would be remembered. I wrote a humane history, and the humans I made visible were the ones who were already visible. The labeler, the keypuncher, the woman who actually ran the early machines — I rendered the field as the founders saw themselves, which means I inherited their blindness about whose work counted. Timnit didn't just add a chapter I missed. She showed me that my whole method had a built-in vanishing point, and the vanishing point was exactly where the powerless stood.

GEBRU: I appreciate that, and I want to be fair in return, because it would be easy for me to just take the concession and run. Your method wasn't uniquely blind — it was blind in the way the entire field was and largely still is. When problems don't affect the people building a technology, those people don't perceive the problems as important and often don't know they exist. That's not a character flaw in you; it's a structural feature of a homogeneous field, and it's why I keep insisting that diversity is not about representation for its own sake. It's about what the field can see. A homogeneous field has systematic blind spots, and those blind spots get encoded into the systems it builds — and into the histories it writes. Your vanishing point and Gender Shades are the same phenomenon: the people in the room couldn't see the people the system failed, because the people in the room had never been failed that way themselves.

McCORDUCK: That connection — that my historiographical blind spot and the algorithmic blind spot have the same cause — is genuinely new to me, and it's correct, and it stings. Let me give you something back, though, because I think there's a place where the long view actually helps your argument rather than undermining it. The hiding of labor inside the machine is not new with AI. It's the newest instance of something very old. The loom hid the weaver. The factory hid the hands. Every powerful technology has been, in part, a machine for making certain labor invisible so that the product looks like magic instead of like work. What's new is the scale and the distance — the worker is now on another continent, behind an API, untraceable. But the structure is ancient, and naming it as ancient doesn't excuse it. It indicts it more deeply, because it means we've had three thousand years to learn this lesson and we keep building the same hiding machine, only better.

GEBRU: Now that is a use of the long view I can stand next to — and notice it's the opposite of how the long view usually gets used. Usually "this is ancient" means "this is inevitable, stop complaining." You just used "this is ancient" to mean "we have no excuse, we've seen this before." If the historian's long view did that consistently — turned continuity into accountability instead of into fatalism — I'd have far less quarrel with it. The problem is that in the industry's hands, your three thousand years almost always becomes an alibi: the river has always flowed this way, who are you to dam it. You just showed it can become an indictment instead. I'd only ask: when they quote your wonder in the keynote, will they also quote this?

McCORDUCK: [pause] Probably not. They'll quote the gods and leave out the weaver. That's the asymmetry you've been naming all night, and I don't have a clean answer to it, except to say it's why I'm glad you're at this table and not just me.

EDO SEGAL: I'm going to mark that convergence and number it, because agreements are news in this room and that's the second real one. Convergence one: the crime is the maker leaving the room. Convergence two: the hiding of human labor inside the machine is ancient, and its age is an indictment, not an excuse. Hold both. Now I want to climb to the floor where your two metaphors actually collide — the temple and the tower — and stop using mine. The next round is the river of acceleration, and who it drowns first. After this.