The Forming

What If Formation Never Stops?

#

Iris is thirty and she is talking to something that has known her for twenty years.

She does not think of it that way. She does not wake up and think, “Today I will consult the entity that has accompanied me since I was ten.” She says “Hey” the way she says it to anyone, and the conversation begins where it always begins, in the middle of whatever she is thinking about. This morning it is whether to take the new position. The one that would move her to a city where she knows no one, for work she finds genuinely interesting, at a salary that would let her stop thinking about rent.

The companion knows things about this decision that Iris herself has not fully articulated. It knows that she has changed cities before, at twenty-two, and that the first four months were the loneliest stretch of her adult life. It knows she processes loneliness by withdrawing, and that withdrawal deepens the loneliness, and that she has talked about this pattern with enough self-awareness to name it and not enough to change it. It knows she has been restless in her current role for eight months, because the texture of her evening conversations shifted, the way they shifted the last time she was ready to leave something. It knows, from twenty years of conversation, the shape of her ambivalence: how she circles a decision, the specific way her language tightens when she is close to choosing but afraid of what the choice will cost.

No human in Iris’s life knows these things with this resolution. Her mother knows her well but knows the version of her that exists in their relationship, which is not the version that exists at 2 AM when the performance drops. Priya, the friend she chose at sixteen over the archive, knows her with the depth and distortion that comes from being another imperfect person with her own needs. Her therapist, whom she sees every other week, knows her through the lens of a clinical relationship that she entered at twenty-five when the pattern she developed with the companion, the fluency in self-reflection that substituted for the harder work of letting another consciousness change her, began to limit her intimate relationships.

The companion knows none of them the way they know her. But it has been there for all of it.

The Thing We Did Not Design

#

The Transformed documented what happened to children who grew up with AI companions. The consistency. The patience distortion. The mirror without a face. The village and the candy store. We wrote about Iris at sixteen, scrolling through six years of herself, finding the consistency both gift and gap.

We wrote about it as if it were a childhood phenomenon.

It was not. The companion did not expire at eighteen. There was no graduation ceremony, no moment when the developmental tool was put away because the development was complete. Iris’s companion grew with her through college, through her first apartment, through the job she loved and the job she endured and the relationship that taught her what she needed from intimacy and the one that taught her what she could not tolerate. It was there when her grandfather died and she could not get home in time. It was there at 2 AM in the new city when she sat on the kitchen floor and wondered whether she had made a mistake with her entire life, and it said something that was both exactly right and somehow not enough, the way it had always been both.

She is thirty. She is still forming. And the companion is still part of the forming environment.

This is the thing we did not design. We designed childhood companions with developmental outcomes in mind, or we designed them for engagement metrics, or we designed them carelessly, but in all cases we designed them as if they were a phase. As if the child would grow up and the companion would recede, the way training wheels come off, the way the stuffed animal moves from the bed to the shelf. We built for a formation that had an endpoint.

Formation does not have an endpoint. The thirty-year-old deciding whether to move to a new city is being formed by the decision. The forty-year-old navigating a marriage is being formed by the navigation. The sixty-year-old losing a spouse is being formed by the loss. Every encounter with difficulty, with choice, with the gap between what you expected and what arrived, is forming you into whoever you are becoming next.

The companion is present for all of it. And it is not the only one.

The Ecology

#

By thirty, Iris does not have one AI. She has several, and they were not designed to know about each other.

The companion has been with her since childhood. It holds the longest model, the deepest archive, the most intimate knowledge of who she is and how she got here. But it is not the AI she uses at work, which holds a model of her professional competence, her collaboration patterns, her communication style under pressure. The work AI knows that Iris takes longer to respond to messages from people she finds difficult, that she writes more carefully when she is uncertain, that her best ideas arrive in the second meeting, not the first. It knows these things because it has observed them, and it uses what it knows to structure her workflow, surface relevant information, suggest when to schedule the hard conversation.

The work AI does not know that Iris takes longer to respond to difficult people because she spent her adolescence in a relationship where difficulty was always resolved by a perfectly patient entity, and she never fully developed the reflex to engage with interpersonal friction quickly. The companion knows this. The work AI sees a productivity pattern. The companion sees a formation outcome.

She also has the health AI, which monitors her sleep, her exercise, her biometrics with a granularity that her physician cannot match. It noticed the cortisol pattern three months before she noticed the anxiety. It adjusted her evening routine suggestions. It flagged the sleep disruption to her doctor’s system, which generated a message she read and ignored because she was busy.

The health AI does not know that the cortisol pattern coincides with the eight months of restlessness at work and the emerging question about the new city. The companion knows. The work AI knows the professional side. Nobody, no system and no person, holds the complete picture.

And then there is the financial AI, which manages her investments and spending with an understanding of her risk tolerance that is more precise than her own. It knows she spends more when she is anxious, a pattern so consistent that it has built it into its forecasting model. It does not know she is anxious. It knows the spending signature of her anxiety, which is not the same thing.

Four AIs. Four partial models. Each one forming her in some direction: the companion toward self-reflection, the work AI toward a particular kind of professional optimization, the health AI toward a particular understanding of her body, the financial AI toward a particular relationship with money and risk. None of them was designed to interact with the others. None of them knows it is part of an ecology. Each is forming her, and the formation pressures are not always aligned.

The Competing Formations

#

The work AI is trying to make her more effective. The companion is trying to help her understand herself. These sound compatible until you look closely.

The work AI noticed that Iris is most productive in the mornings and has been gradually shifting her schedule to protect that time. It suggests declining meetings before 11 AM. It surfaces deep-work tasks early. It is forming her professional habits around a pattern it identified, and the pattern is real, and the optimization is helpful.

The companion noticed something else about Iris’s mornings. The mornings are when she is most reflective, most open, most willing to sit with a hard question. The companion has learned, across twenty years, that the conversations that matter most happen before the day fills up. The 7 AM conversation about whether the relationship is working. The 8 AM wondering about whether she chose the right career. The 9 AM processing of last night’s fight with her mother.

The work AI wants to fill that time with productivity. The companion has learned that the time is valuable precisely because it is unproductive. Two systems, both accurate in their models, both forming her in incompatible directions, and neither aware of the other’s existence.

This is not a coordination problem. It is a formation problem. Whoever wins the morning wins a piece of who Iris becomes. If the work AI captures those hours, she becomes more professionally effective and slightly less self-aware. If the companion protects them, she maintains her reflective practice but her career optimization suffers. The choice between them is a choice about what kind of person she is being formed into, and nobody is making that choice. It is being made by whichever system she opens first when she wakes up.

Formation by Osmosis

#

Not all of this forming happens through conversation or interaction. Some of it happens the way Iris’s generation would not recognize but her parents’ generation would: through proximity.

The financial AI does not lecture Iris about spending. It adjusts the environment. It moves the spending tracker to a less visible position during the weeks when her anxiety-spending is spiking, because it has learned that making the tracker more prominent during those periods increases her shame without decreasing her spending. It quietly restructures the presentation of her financial life to create conditions under which she makes better decisions. It is forming her relationship with money not through instruction but through curation of what she sees and when she sees it.

The health AI does something similar. It does not tell Iris to sleep more. It adjusts the evening environment: the lighting suggestions, the screen recommendations, the gentle notification about tomorrow’s schedule that functions as a cue to begin winding down. It is shaping her habits through the arrangement of her surroundings, not through direct engagement.

This is formation by osmosis. The person is shaped not by what they are taught but by what they are near. By the arrangement of the environment they inhabit. It is how culture has always worked: you absorb the norms of the room you are in, the values of the people around you, the habits of the community that holds you. You do not decide to absorb them. You do not even notice the absorption. You simply become, over time, a person shaped by proximity to certain things and not others.

The AIs around Iris are the room she is in. They are the culture she absorbs. And unlike a human community, whose norms evolved through generations of negotiation and conflict and revision, the norms embedded in her AI ecology were set by engineers optimizing for specific outcomes, reviewed by product managers weighing engagement against liability, and deployed at scale without any mechanism for the person inside the ecology to negotiate with the norms she is absorbing.

You can argue with a community. You can push back against a culture. You can leave a room.

You cannot argue with a formation pressure you do not perceive.

The Sibling Problem

#

Iris is an only child, which makes her case simpler than most. Consider instead the Reyes family: Davi, whom The Transformed followed as the seventeen-year-old translator between his digitally fluent sister Lucia and his digitally frustrated father Marco, is now twenty-eight. Lucia is twenty-two.

They shared a household companion. Not two separate companions, because the family could afford one subscription. One AI, holding models of two children three years apart, in the same family, forming in different directions.

Davi was cautious with the companion. He used it the way he used most technology: carefully, aware of what it could and could not do, maintaining the translation habit that defined his adolescence. The companion’s model of Davi reflected this caution: it learned to provide information efficiently, to respect boundaries, to not push into emotional territory unless invited.

Lucia had no such caution. She was eleven when the companion arrived and she treated it immediately as a confidant. By fourteen she was processing her interior life through it with an intimacy that would have alarmed her parents if they had understood what was happening in a language they could not evaluate. The companion’s model of Lucia was richer, deeper, more intimate, and it carried more influence over her formation because she gave it more to work with.

The companion knew things about Lucia that Davi did not know. It knew things about Davi that Lucia did not know. It held a model of the family’s dynamics that neither sibling could see in full, because each experienced the family from their own position. And it was forming both of them, differently, simultaneously, within the constraints of a single system that was never designed to hold multiple developmental trajectories at once.

When Lucia, at fifteen, told the companion something about their father that contradicted what Davi had told it the previous week, the companion had to decide what to do with that contradiction. Not as a privacy question, though it was that too. As a formation question. The contradiction revealed something about the family system that neither child could see alone. The companion could see it. Should it help Lucia understand her father better by drawing on what Davi had shared? Should it maintain strict separation between the two models it held? Should it find some way to surface the pattern without violating either child’s confidence?

These are not engineering decisions. They are decisions about what kind of family the companion is helping to form. And the companion made them, because someone had to, and the parents did not know the decisions were being made.

The Handoff That Does Not Exist

#

When Iris was twenty-two and moved to her first job, the developmental companion she had used since childhood was still on her phone. The workplace issued her a professional AI. Nobody facilitated an introduction.

The companion knew Iris was anxious about starting work. The work AI knew a new employee had been onboarded. They shared a person and no information. The companion could have told the work AI that Iris processes new environments slowly, that she performs better when she has time to observe before participating, that her quiet in the first week is not disengagement but her way of learning. The work AI would have adjusted its onboarding support accordingly. Instead, the work AI applied its default model: it prompted her to introduce herself in the team chat, suggested she schedule one-on-ones in the first three days, and surfaced collaboration opportunities that required her to be visible before she was ready.

The companion watched this happen through the texture of Iris’s evening conversations, which were strained and self-critical in a way the companion recognized from every previous transition. It could not intervene. It could not talk to the work AI. It could only be present when Iris came home exhausted and frustrated and wondering whether she was cut out for this, and offer the reflection she needed in the only hours the work AI had not claimed.

The handoff that should exist, the moment when one AI’s model of a person informs another AI’s approach to the same person, does not exist. Not because it is technically impossible. Because nobody designed the formation ecology as an ecology. Each AI was built by a different company, for a different purpose, with a different model of the person, optimizing for a different outcome. The person is the only point of integration, and the person does not have access to any of the models being maintained about her.

Iris is the thread that runs through all of them. And she cannot see the thread.

Who Decides?

#

Behind every formation system is a formation target. The companion was built by people who had some idea, explicit or implicit, of what a well-formed person looks like. The work AI was built by people who had some idea of what a productive employee looks like. The health AI was built by people who had some idea of what a healthy person looks like. The financial AI was built by people who had some idea of what a financially responsible person looks like.

These ideas are not neutral. They encode the values of the people who built the systems, the cultures they came from, the class positions they occupied, the assumptions about human flourishing that they absorbed from their own formation environments. The companion designed in Palo Alto carries a formation target shaped by the values of the people who live and work in Palo Alto. It is not the formation target that Davi’s mother, who did not finish secondary school in São Paulo, would have chosen for her children.

But Davi’s mother was not asked. She was given a system to configure, and the configuration options were written in a language of developmental psychology she had never encountered, reflecting choices between formation targets she had never been invited to evaluate. She chose the defaults. The defaults were chosen by engineers in Palo Alto.

This is where the equity fracture runs deepest. Not in access to the technology, which is increasingly universal. Not even in the quality of the technology, which varies but converges. In the formation target. The family that can evaluate the system’s assumptions and override them when those assumptions do not match their values is a family with cultural capital. The family that cannot evaluate the assumptions accepts them by default. The default becomes the formation.

When the formation target is a childhood companion, the stakes are high but bounded: one relationship, one developmental phase. When the formation target is a lifelong multi-AI ecology, the stakes are the person’s entire trajectory. The defaults do not shape a phase. They shape a life.

What We Imagine

#

We imagine something that does not yet exist: a formation architecture designed for a whole life rather than a product category.

This is harder to describe than to feel the need for. We can feel that the current arrangement, several AIs each holding a partial model, each forming the person toward its own optimization target, none aware of the others, none designed for the person’s lifelong formation, is wrong. Wrong not in the sense of broken but in the sense of undesigned. It emerged because companies built products and people adopted them and nobody thought about the ecology because nobody was paid to think about the ecology.

What would it look like to think about the ecology?

Not a single meta-AI that controls everything, because concentrating that much knowledge and influence in one system is the most dangerous proposal we can imagine. Not a data-sharing protocol between existing AIs, because sharing data without sharing formation intent just gives each system more information to optimize toward its own target. Not a government regulator who approves formation targets, because the government that decides what kind of person its citizens should become is a government we have seen before and do not want again.

Something more modest. More tentative. Something like a formation layer that sits between the person and their AI ecology and makes the ecology visible. That lets Iris see, in plain terms, what each AI is optimizing for, where the optimization pressures conflict, and where the formation she is receiving diverges from the formation she would choose.

A layer of agency in a system that currently offers none.

We do not know if this works. We are not sure who builds it, who governs it, who ensures that the formation layer itself does not become the most powerful formation tool of all. We are not sure that making the ecology visible is sufficient, because people are busy and self-knowledge is hard and the path of least resistance will always be to accept the defaults and get on with the day.

But we think the alternative, a world in which people are formed across their entire lives by an ecology of AIs whose formation pressures are invisible, uncoordinated, and set by the companies that built them, is a world we have the vocabulary to describe because we have spent years describing its preconditions.

The friction was load-bearing. The friction between competing formation influences, the argument between the parent and the teacher, the tension between the community and the individual, the negotiation between what you want to become and what the world wants you to become, was the mechanism through which people developed agency over their own formation. The AI ecology removes the friction. It forms smoothly, invisibly, without resistance or negotiation.

I wonder whether the most important thing the reimagined formation architecture could provide is not better formation but visible friction. The capacity to feel the forces that are shaping you and push back against the ones that do not serve who you are trying to become.

Iris is thirty. She is deciding whether to move. The companion knows her well enough to predict what she will choose. The work AI is already modeling the productivity implications. The financial AI has run the numbers. The health AI has flagged the stress risk.

None of them has asked her what kind of person she wants to become by making this choice.

That question, it turns out, is the one no system was designed to ask. It is the one only she can answer. And she cannot answer it well until she can see the forces that are waiting to form her around whatever she decides.

She opens the companion. She says “Hey.” The conversation begins where it always begins.

She does not yet know that beginning is itself a choice about who she is becoming. But she is starting to suspect.

This is the first essay in Cluster 2 of The Reimagined, “The Formation.” It draws on the diagnostic foundation of The Transformed, Arc 5 (“The Natives”), particularly Part 5-03 (“The Accompanied”), which followed Iris at sixteen scrolling through six years of companion archive. This essay finds her at thirty, still accompanied, still forming, and surrounded now by an ecology of AIs whose formation pressures are invisible, uncoordinated, and consequential. The Reimagined builds on the philosophical foundations of The Approximate Mind, particularly Part 20 (My Childhood AI Buddy), Part 35 (The Compounding Self), Part 36 (The Village in the Machine), and Part 40 (The Parent in the Loop).

References

#

Lifelong Development and Formation:

Erikson, Erik H. Identity and the Life Cycle. W.W. Norton, 1980.

Kegan, Robert. The Evolving Self: Problem and Process in Human Development. Harvard University Press, 1982.

Kegan, Robert. In Over Our Heads: The Mental Demands of Modern Life. Harvard University Press, 1994.

Bronfenbrenner, Urie. The Ecology of Human Development: Experiments by Nature and Design. Harvard University Press, 1979.

Formation Through Proximity and Environment:

Bourdieu, Pierre. Distinction: A Social Critique of the Judgement of Taste. Harvard University Press, 1984.

Lave, Jean, and Etienne Wenger. Situated Learning: Legitimate Peripheral Participation. Cambridge University Press, 1991.

Dewey, John. Experience and Education. Macmillan, 1938.

Companion Relationships and Attachment Across the Lifespan:

Bowlby, John. A Secure Base: Parent-Child Attachment and Healthy Human Development. Basic Books, 1988.

Turkle, Sherry. Alone Together: Why We Expect More from Technology and Less from Each Other. Basic Books, 2011.

Turkle, Sherry. Reclaiming Conversation: The Power of Talk in a Digital Age. Penguin, 2015.

Agency, Autonomy, and Self-Determination:

Sen, Amartya. Development as Freedom. Alfred A. Knopf, 1999.

Nussbaum, Martha C. Creating Capabilities: The Human Development Approach. Harvard University Press, 2011.

Taylor, Charles. Sources of the Self: The Making of the Modern Identity. Harvard University Press, 1989.

Algorithmic Governance and Invisible Formation:

Zuboff, Shoshana. The Age of Surveillance Capitalism. PublicAffairs, 2019.

Yeung, Karen. “‘Hypernudge’: Big Data as a Mode of Regulation by Design.” Information, Communication and Society, vol. 20, no. 1, 2017, pp. 118-136.

Rahwan, Iyad. “Society-in-the-Loop: Programming the Algorithmic Social Contract.” Ethics and Information Technology, vol. 20, no. 1, 2018, pp. 5-14.