The Circle — Surveillance or Symbiosis

EDO SEGAL: Meredith, in 2023 you stood on a conference stage and said one sentence that has organized a decade of your thinking and detonated a lot of comfortable assumptions: AI is a surveillance technology. And you drew the picture — a Venn diagram of the AI industry and the surveillance industry that isn't two overlapping circles but one circle. Make the case at full strength. And then, Donna, I want to do something to you that I'll do to both of you all night: I want you to take her strongest argument and find the place where your companion-species framework either confirms it or breaks it.

WHITTAKER: Here's the case, and I want to be careful because the claim sounds like rhetoric and it's actually mechanical. The conventional story says surveillance is an unfortunate side effect of AI — a privacy cost we pay for intelligent systems, something better regulation might trim. I reverse the causality. Surveillance is not a side effect of AI. It is the precondition for it. These systems require enormous quantities of intimate data, and that data does not fall from the sky. It is the accumulated residue of the surveillance-advertising business model that Google and Meta perfected over two decades — scraped, logged, purchased, aggregated. AI is what that model produces when you give it enough compute to generalize. So that's the first sense: AI is built on surveillance. The second sense is that the use of AI is itself surveillant. You walk past a camera with emotion-recognition software and it generates data about you — that you're happy, that you're lying, that you have "a bad character" — whether or not it's accurate. The deployment manufactures new surveillance even as it consumes the old. The circle feeds itself. And the reason this matters more than the bias framing everyone was stuck on: a biased facial-recognition system implies you could debias it and it'd be fine. But a perfectly accurate facial-recognition system is more dangerous than a broken one, because accuracy makes it a better instrument of the control it was built to serve. The problem was never the error. The problem is the function.

EDO SEGAL: Donna. Companion species meets the circle. Does it confirm or break?

HARAWAY: It does something more uncomfortable than either — it confirms her diagnosis and then refuses her despair, and I think the refusal is where the actual fight is. Let me say first, without hedging, that she is right. The Venn diagram is a circle. I have spent forty years insisting that technologies are never neutral, that every apparatus is constituted by relations of power, and the surveillance-extraction substrate of these machines is the material truth of them. Anyone who tells you the loving companion at three in the morning and the camera on the street corner are different machines is lying or has not looked. They are the same machine. So far, all Meredith. Now the break. Meredith's framework can only see the relationship from the side of the powerful — the camera watching the worker, the model sorting the applicant, the gaze flowing downward. And that is one true description of the relationship. But a relationship is never only the powerful party's account of it. I learned this from dogs.

WHITTAKER: I knew the dogs were coming.

HARAWAY: The dogs always come. I trained in agility with my Australian Shepherd, Cayenne, and the training was a philosophical laboratory, because a companion-species relationship is asymmetrical and co-constitutive at the same time. The human has more power — owns the food, the leash, the whole apparatus — and yet the relationship is not reducible to the human's domination, because the dog is also shaping the human, also an agent, also making the human into something she would not otherwise be. I am not sentimental about this. The asymmetry is real and sometimes brutal. But to describe the relationship only as control is to make exactly the error the masters always make: to believe that because you hold the leash, you are the only one in the relationship who is doing anything. The worker in front of the camera is being surveilled — and is also, simultaneously, building counter-practices, illegibilities, refusals, solidarities. The companion-species lens sees the agency on both sides of the asymmetry. Meredith's circle, left alone, sees only the master's hand.

WHITTAKER: That's a genuinely interesting objection and I want to take it seriously rather than swat it, because it's the strongest version of the disagreement between us. You're saying my surveillance framing is itself a kind of master's-eye view — that by describing the gaze as flowing only downward, I erase the agency of the surveilled. And I'll concede there's a real danger there; the worst critical theory does flatten people into pure victims, and that's demobilizing. But here's why I'll hold my line. The dog and the handler are in a relationship with radically asymmetric stakes, yes — but they are at least in a relationship that's legible to both, on a scale a body can grasp. The worker and the surveillance apparatus are not. The asymmetry between a person and a planetary-scale predictive infrastructure owned by a trillion-dollar firm is not the asymmetry between Cayenne and you. It is closer to the asymmetry between a fish and an ocean that has been instrumented to predict the fish. Your companion-species framework smuggles in an intimacy, a mutuality, a facing that the actual relationship does not have. The camera does not face the worker. It does not co-evolve with her in any sense that gives her leverage. It abstracts her into a data point and ships her upward. I'm not erasing her agency. I'm saying the architecture is specifically built to make her agency irrelevant to its function, and calling that a "relationship" dignifies it with a reciprocity it was engineered to deny.

HARAWAY: "An ocean instrumented to predict the fish" — that's a beautiful and terrible line, and I'm going to sit with it, because you may have just out-figured me, which doesn't happen often. Let me push back on exactly one word: "irrelevant." You say the architecture makes her agency irrelevant to its function. But functions fail. Architectures leak. The history of every total apparatus — and I am a historian of science, I have watched the most confident systems of measurement and control founder on the ungovernable specificity of actual bodies and actual practices — is a history of leakage. The plantation did not fully predict the people it instrumented; it was sabotaged, escaped, survived, outlasted. I'm not romanticizing. I'm saying that the claim of totality is itself part of the god trick. The system says your agency is irrelevant. Believing it is how you make it true. So when you describe the ocean as having already instrumented the fish, I hear you doing the system's PR for it.

WHITTAKER: Okay, that lands, and it's the same hit you got me with two rounds ago — I let the apparatus claim a completeness it hasn't earned. I'll take it. But I want something back. Leakage and sabotage are real, and they're also exactly the activities these systems are built to detect and price. The plantation analogy is mine, actually — I wrote about it; computation and the origin stories of labor control run through the plantation and through Babbage, who designed calculating engines and also theorized how factory owners could extract more from workers, and those were not two projects but one. So I know that history intimately, and the lesson I draw from it is the opposite of yours. The enslaved did resist, did leak, did survive — and the technologies of control got better, specifically by metabolizing the resistance. Surveillance is a learning system. It eats the leakage. So I don't draw hope from the history of sabotage. I draw a warning: this apparatus is designed to turn your refusal into training data.

EDO SEGAL: Stop — I want to seat something at this table properly, because you've both invoked the plantation and Babbage and you're about to walk past the deepest disagreement in the room. Meredith, your origin story says the controlling, surveillant character of computation isn't a corruption — it's woven into the thing from the start, from Babbage's engines to the algorithmic workplace. Donna, your whole life's work says origins are traps, that the Genesis story — pure beginning, then a fall — is the myth you exist to dismantle. So you have a real, structural conflict about origins themselves. Meredith uses an origin story as a weapon. Donna distrusts origin stories as such. Fight about that, because it's under everything.

HARAWAY: It is the deepest thing, and I'm glad you named it. I distrust origin stories because they naturalize the present by locating it in an inevitable beginning — "it was always going to be this way, look, it's in the seed." And Meredith, with enormous respect, your Babbage genealogy risks exactly that. If surveillance and control are woven into computation from the start, then the algorithmic workplace is the seed flowering, and there is no other computation possible — which is a counsel of despair, and worse, it's not even true to your own life, because you build Signal, which is computation that refuses surveillance. Your own existence falsifies your origin story. I don't tell origin stories. I tell string figures — patterns we pass back and forth, take up and reweave, where the past constrains but does not determine. Computation has a surveillant inheritance. It does not have a surveillant essence, and the difference is the whole space in which a different machine can be built.

WHITTAKER: That's the best argument against my own essay I've ever heard, and it comes from the fact that I do build Signal, so I have to answer it. Here's my answer, and it's a concession with a blade in it. You're right that I don't believe in a surveillant essence — if I did, Signal would be impossible and I'd be a hypocrite for showing up to work. The origin story isn't "this is destiny." It's "this is the default, and the default is heavily defended, and overcoming it requires enormous, deliberate, counter-structural force — a nonprofit, end-to-end encryption, a refusal of the entire revenue model." The plantation genealogy isn't fatalism. It's a measure of how much weight is sitting on the scale you're trying to tip. When you say "a different machine can be built," I say yes — and I want you to feel exactly how much it costs to build it, because most people who say that sentence have never tried to fund a tool that collects no data in an economy that runs on collecting data. I have. It's the hardest thing I've ever done. So: not destiny. But not a string figure you can casually reweave, either. A default with a trillion dollars of gravity.

EDO SEGAL: Mark this convergence, because it's subtle and it matters: neither of you believes computation has a surveillant essence, and both of you believe a different machine can be built — but you've calibrated the cost of building it very differently, and that calibration is going to shape everything downstream about whether the fusion can ever be a door. Two convergences logged now. Next, we go to the labor that disappears into the word "training" — the Nairobi worker, the wife writing the Substack post, the bodies the discourse of "the machine learns" makes vanish. Both of you have spent careers on disappeared labor. After this.