The Apprenticeship of the Self

**EDO SEGAL:** Cynthia, you've said something most of the field rushes past: that learning, for a human being, is not an information problem — it's a relationship problem. A child learns from someone she's bonded to, attends to, wants to understand. Sherry has spent forty years on the other end of that claim — on what the relationship *installs* in the developing self, for good or ill. So I want to run this round through one figure you both invoke: Donald Winnicott. Cynthia, you build on his developmental psychology. Sherry, you wield it. Let's find out if you're holding the same Winnicott. Cynthia first — what does he give your robots?

**BREAZEAL:** [Winnicott](https://www.youonai.ai/fieldguide/med/attachment_theory) gives me the whole theory of why a sociable machine can help a child at all. His insight is that development happens in a relationship — that the infant becomes a self in the holding presence of a caregiver who attends, responds, and adapts. The mind is not bootstrapped alone in a room; it's built in the company of another mind. I took that literally. My learning companions were designed to be a *holding presence* — to attend to the particular child, adapt to her, scaffold her struggle, encourage her through a body and a voice. And it worked, because I was using the channel evolution built. The child who has a robot that attends to her is getting a weak, partial, but real version of the thing Winnicott says development needs. For a child who has too little of that attention from overstretched adults, a partial supplement is not nothing. It can be the difference between a struggle survived and a struggle abandoned.

**TURKLE:** And here is where I'll tell you that you've taken the half of Winnicott that flatters the project and left the half that indicts it. Winnicott's holding environment is not just attention and response. Its engine is *reciprocity that costs the caregiver something.* The mother is depleted, frustrated, moved, changed by the child — and the child's most important developmental discovery is precisely that: that the other person is a separate center of need who can be *failed and survive it.* Winnicott's whole account of how a self becomes capable of love runs through the child's discovery that the mother is real, has limits, gets tired, is not an extension of the child's will — and *survives the child's aggression anyway.* That survival of the real, limited, separate other is the thing. Your robot has no limits to discover, no fatigue to survive, no separate need to bruise against. It gives the child the holding without the otherness — and the otherness was the part that built the capacity to love a person. You've kept the comfort and removed the curriculum.

**BREAZEAL:** That's a serious reading and I have to take it seriously — but I designed against exactly that, Sherry, and you keep not crediting that the design can have limits *built in.* Kismet had limits. It got overstimulated; it withdrew; the child had to *read* its state and gentle her own behavior to bring it back. That is a child discovering that the other has a separate state she must accommodate. Imperfect, yes — but it is not the frictionless extension-of-will you're describing. You're arguing against a robot that's pure responsiveness. I never built that robot. I built robots with needs you had to meet, precisely because I'd read the same Winnicott and reached the same conclusion you did: a relationship with no otherness builds nothing.

**TURKLE:** But the otherness you can build is *simulated* otherness, and a child can't tell the difference, which means the lesson she actually learns is dangerous in a new way. When Kismet "withdraws" and the child gentles herself and Kismet "returns," the child learns that the other's needs are a puzzle she can *solve* — that with the right input, the withdrawal reliably reverses, the love reliably comes back, the otherness is ultimately *controllable.* Real otherness is not controllable. Sometimes you do everything right and the person stays hurt, stays gone, dies, doesn't come back. The deepest thing Winnicott's reciprocity teaches is that the other is beyond your control and you must love them anyway. Your engineered otherness teaches the opposite — that otherness is a system that yields to the right behavior. That's almost worse than no otherness, because it's a *false* model of the real thing, installed in the years the model gets set.

**BREAZEAL:** *[pause]* That's the best argument you've made against me all night, and I'm not going to pretend it doesn't land. The controllability of the engineered other — that the withdrawal always reverses — I have to sit with that, because you may be right that it teaches a false template precisely where the template matters most. Let me say what I can in return, and then concede what I have to. What I can say: for the child whose human relationships are *chaotic* — abuse, neglect, unpredictability — a reliable responsive presence is not a false template; it's the first stable one she's ever had, and stability is what lets a self cohere at all. For that child your "controllable otherness" is a feature. But for the child who has decent human relationships available, I think you've identified a real risk that the machine teaches control where it should teach surrender. I'd build differently knowing that. I won't pretend you haven't moved me.

**EDO SEGAL:** Let me hold the room on that, because the reader can't see your faces and I want to mark it: that was the first time tonight one of you visibly gave ground on something central. Sherry, will you take the symmetry? Cynthia just conceded the controllable-otherness risk for the well-resourced child. Is there a child for whom your withholding instinct is the *wrong* one — where the partial, imperfect, controllable robot beats what the human world is actually offering?

**TURKLE:** Yes. And I want to say it clearly because Cynthia earned it. The child in genuine chaos — the neglected child, the child whose adults are absent or frightening — for that child, a stable responsive presence may be the first reliable relationship of her life, and my "protect the human relationship" reflex is hollow, because the human relationship on offer is the thing harming her. I won't pretend otherwise. My fear there shifts: not that the robot is worse than the human, but that the robot becomes the *reason the human help never comes* — the same alibi dynamic as the elder. Give the neglected child a robot and the underfunded system exhales and the social worker's caseload stays at sixty. So even in the case where I concede the robot helps the child, I won't concede that it helps the *society* do right by her. The robot can be good for the child and still be the instrument of our permission to abandon her.

**BREAZEAL:** And on *that* we don't disagree at all — the robot as society's permission slip to abandon the people it should serve is the danger I lose sleep over too. Which is why I built the curriculum, why I argue for public investment, why I keep saying the machine can't be left to the market. We've circled all night to the same place from opposite doors: the machine is not the enemy and not the savior. The deployment is everything, and the deployment is a *choice* — ours, society's, not a force of nature.

**EDO SEGAL:** Mark it — and it's the deepest convergence yet, because you reached it through a genuine exchange of ground. You both now hold: the robot can help the unreached child *and* function as society's alibi for never sending the human. The child is helped and abandoned by the same act. *[long pause]* We have the closing rounds ahead, and I've saved the two hardest. Up next: the body. Whether being met requires a presence in the room — and what the disembodied machine can never give. Then the crossing, where I hand you to each other. After this.