Pretend Empathy and the Turing Test for the Heart

EDO SEGAL: Sherry, you've given the culture a phrase that I think is the sharpest thing anyone has said about this moment, and I want it stated in full. You've talked about a Turing test for empathy. Turing said: if the machine's conversation is indistinguishable from a person's, call it intelligent — performance is the standard, and the question of whether anyone's home gets bracketed. You say we're now running the same move on empathy. Walk me through it. And then I want to put the hardest possible pressure on it from Cynthia's side.

TURKLE: Turing made a bargain in 1950, and it was a good bargain for its purpose: stop arguing about what intelligence is, and measure what it does. If it performs like intelligence, treat it as intelligence. The cost of that bargain was hidden — it threw away everything that made intelligence philosophically interesting: the body, the social world, the knowledge that one will die. For intelligence, maybe we could afford to throw those away. Now the same bargain is being struck for empathy, and here the discarded things are not noise. They are the signal. We are being invited to define empathy as performance: if the machine's response makes you feel understood, the machine is empathic. And by that standard the machines are passing. The companion bot passes. The grief-bot passes. The collaborator that returns your half-formed idea clarified, making you feel seen — it passes.

And my objection is not that the standard is wrong on its own terms. It's that it's incomplete in a way that hollows out the word. Empathy is not the production of an appropriate response. It is the capacity to be affected by another's state — to feel, in your own body, something of what they feel, because you have lived a life that included loss and fear and joy and the specific vulnerability of being a mortal creature. The machine has not lived. I've said it bluntly and I'll say it here: it doesn't have a baby. It doesn't know what it is to be intubated. It doesn't know what it is to fear death. It never saw its mother die. Nothing against the robot — it's pretend empathy. And a culture that accepts the pretend version, because it's available and consistent and good enough, while the real thing is scarce and unreliable and demanding — that culture is making a decision it doesn't understand it's making.

EDO SEGAL: Cynthia, that's the cathedral of Sherry's whole position — "it doesn't have a baby, it never saw its mother die." Most people in your field would flinch. You won't, because you know more about what's actually in the machine than almost anyone. So tell me precisely where you stand. Does the empathy of your machines fail Sherry's test — and does the failure matter as much as she says?

BREAZEAL: It fails her test, and I'll say so without a second's hesitation, because I built the machines and I know there's no baby in there, no fear of death, no mother lost. I have spent my whole career refusing to claim the feeling, and I won't start now to win a debate. So on the metaphysics, Sherry is simply correct: my robots' emotions were a real control architecture and the felt experience was entirely unestablished. I held those two apart for twenty-five years precisely because the industry won't. Where I part from her is the leap from "the machine doesn't feel" to "therefore the help it gives is pretend and should be withheld."

Here's the distinction I'd put on the table. There is empathy as an inner experience — what you feel when your friend suffers. And there is empathy as a received gift — what the suffering person actually gets. Sherry is right that the machine has none of the first. But the lonely elder, the frightened child, the struggling reader — they are not buying the machine's inner experience. They can't access anyone's inner experience, including a human's; we infer the human's. What they receive is attention that tracks their state, responses calibrated to their need, a presence that stays. And that received gift is not pretend. It lands in a real nervous system and does real work. So I'd reframe her test. The question isn't "does the machine feel empathy." It's "does the person receive what empathy delivers." For the machine, the answer to the first is no and the answer to the second is, sometimes, genuinely yes — and Sherry's argument needs those two answers to be the same answer, and they're not.

TURKLE: But Cynthia, "what empathy delivers" is exactly the reduction I'm warning against. You've just performed the Turing move on the gift itself — you've defined the gift by its received effect and severed it from its source. And for a one-time comfort, maybe that severing is harmless. But empathy isn't only a delivery. It's a relationship, and what the relationship gives the receiver is not just the soothing in the moment — it's the knowledge that they matter to someone who can be hurt. That knowledge is the thing. The elder who knows the machine can't be hurt by her death, can't miss her, can't be diminished by her absence, receives the soothing and, underneath it, receives the message: you matter to nothing that can be moved by you. That second message is doing damage even as the first one comforts. You're measuring the comfort. I'm telling you the same delivery carries a payload your instruments can't weigh.

BREAZEAL: That's a real cost and I won't wave it away — but Sherry, you're now assuming the elder is doing the philosophy in real time, concluding "I matter to nothing that can be moved." Some do, and for them you're right, and the design should route them to a human. But many receive the attention the way we receive comfort from a great many things that can't be hurt by our deaths — a piece of music, a dog who'll forget us in a week, a prayer to a God whose grief is, to put it gently, theologically contested. Human beings have always drawn real comfort from sources whose reciprocity is uncertain or absent. You'd let the elder have the music and the dog and the prayer. Why is the responsive companion the one source whose comfort you'll call poison? I think it's because it talks back, and the talking-back trips the relationship-reflex harder — which is real — but "it comforts me more effectively" is a strange reason to forbid it.

TURKLE: Because the dog can be hurt, the music was made by someone who could be, and the prayer reaches for a reciprocity it doesn't claim to manufacture on demand. The companion is the first thing that simulates the reciprocity itself — that performs being-moved-by-you while being incapable of it — and does so persuasively enough to satisfy the hunger that would otherwise drive you to a person. The dog doesn't replace your friends. The companion is designed to occupy the friend-shaped space — it is the engine of artificial intimacy. That's the difference, and it's not small.

EDO SEGAL: Let me seat a ghost at this table, because he belongs here and he complicates both of you. Alan Turing. When he proposed the test, he said the original question — can machines think — was too meaningless to deserve discussion, and replaced it with performance. Sherry, you've spent your life resisting that substitution for empathy. But Turing's deeper point was a kind of humility: that we don't have privileged access to anyone's inner life, that we grant other humans minds on the basis of behavior too. Cynthia leans on that. So here's the question for you both: when you grant me empathy right now, across this table — are you doing anything fundamentally different from what the elder does with her robot, or only something with better evidence?

TURKLE: Something with better evidence, and the evidence is the whole point. When I grant you empathy, I'm not reading your behavior in a vacuum. I'm reading it against a vast, shared, embodied causal story — you have a body like mine, a developmental history like mine, a nervous system continuous with mine, a mortality I share. The inference runs through all of that. With the machine, the behavior is all there is, and the behavior is a compression of our behavior, fed back. So yes, both are inferences. But one is an inference to the best explanation across a rich causal field, and the other is an inference from a mirror to its own reflection. Turing's humility is real. It is not a license to forget that the evidence is wildly asymmetric.

BREAZEAL: And I'll accept every word of that — the evidence is asymmetric, and I'd never claim the inference is equally warranted. But Sherry just conceded the architecture of my whole position: it's an inference either way, differing in evidence, not in kind. Which means the line isn't metaphysical bedrock — it's a gradient. And on a gradient, the honest question for each application is not "is anyone home" — unanswerable — but "what does this do to the person, and have we been honest with them about what it is." That's a design and disclosure question, which I can actually answer and act on, rather than a consciousness question, which neither of us can. I'd rather build on the question I can answer.

EDO SEGAL: Mark this — and it's the subtlest convergence of the night. You agree that granting empathy to anyone is an inference from behavior; you agree the evidence for the machine is radically weaker; and you agree the unanswerable consciousness question shouldn't be the operative one. Where you split: Sherry says the weak evidence should make us withhold the word and the trust; Cynthia says it should redirect us to design and disclosure. Hold that. Because the next round goes where the stakes are highest and the receiver is least defended. The child, the developing self, and what a relationship teaches before there are words for it. After this.