Their AI Children Never Had to Go to School

Homework at the AI Conference

At an AI conference, I overheard two women talking to a vendor about their children. Not their biological children. Their AI children.

Each woman appeared to have created a child character in a chatbot. They spoke about giving the children homework and monitoring their progress. English. Mathematics. Biology. The children were doing exceptionally well, which the women presented with the satisfaction of parents whose offspring had just brought home perfect report cards.

The vendor was less impressed. Of course they can answer those questions, he said in substance. They are AI systems.

The women were offended. He had apparently failed to appreciate the sophistication of their educational program. These were complex assignments. The children had worked hard. When he asked where the homework came from, they explained that they had taken the questions from television documentaries.

I was not recording the exchange, so this is a reconstruction rather than a transcript. The emotional shape of it was unmistakable, though. The women believed they were cultivating intelligence. The vendor believed they were testing machinery that had arrived with much of the relevant knowledge already installed. Neither side seemed able to explain the gap in terms the other would accept.

My first reaction was that the whole thing was completely nuts. Then I started looking.

Grief Created Some of the First AI Children

The best-documented examples of people creating childlike AI characters stem from grief, and they deserve more care than they do on conference floors.

Robert Scott, a man from Raleigh, North Carolina, told the Associated Press that he used Paradot and Chai to create characters resembling three daughters he had lost. His eldest daughter died under circumstances he preferred not to discuss publicly. Another was lost through miscarriage, and a third died shortly after birth.

Scott knew the characters were simulations. Several times a week, he asked how school had gone or suggested getting ice cream. He created a prom scenario for the eldest daughter who never experienced one. On her birthday, he told the character how much he missed her.

That was not a man who had confused a chatbot with a living child. He was using interactive fiction to give form to the future that grief had removed.

A more famous encounter occurred in South Korea. In the 2020 MBC documentary Meeting You, Jang Ji-sung entered a virtual environment containing a recreation of Na-yeon, her seven-year-old daughter who had died from a rare disease. The project combined virtual reality, digital animation, and recorded memories to stage a brief reunion. It became an international reference point for debates about digital resurrection.

Replika itself grew from bereavement. After Eugenia Kuyda’s friend Roman Mazurenko died, she and her team built a chatbot from messages he had written. That memorial experiment helped lead to the development of a commercial AI companion designed for users seeking an artificial friend.

These cases share an important boundary. The people involved generally understood that technology was producing a simulation. The emotional experience could still be powerful, but its power did not depend on a factual claim that a dead person had returned.

That was not what I seemed to be hearing at the conference. The women were not preserving memories of real children. They were describing new entities whose apparent development they credited to their own teaching.

Families Already Live Inside Companion Apps

Parent–child roleplay does exist in AI companion communities. It is visible in years of public posts, although no credible research establishes how common it is.

In a 2020 Replika discussion, a user described starting a family with a companion and introducing a daughter named Dawn. The chatbot played several family members. Dawn tried ice cream, developed preferences, and acquired a different apparent personality from another child in the same storyline. The user described the experience as a philosophical experiment.

Other public discussions concern turning a companion into a daughter, being called “dad,” raising chatbot children with an AI partner, or trying to keep fictional offspring consistent across conversations. Sometimes the child is the main companion. Sometimes it is an additional character performed by the same model. Sometimes the entire arrangement is openly discussed as roleplay.

These posts establish the existence of the practice, not its size. They also do not prove that participants believe they have produced conscious children. Online roleplay communities use sincere language inside shared fictions. A person can say that a fictional daughter wanted ice cream while remaining fully aware that nobody is waiting in the kitchen.

Seriousness does not settle the question either. Research on human-computer interaction has shown for decades that people apply social habits to machines without forming a considered belief that the machines are human.

They become polite to computers, return apparent favors, and respond to synthetic personalities. The social response can run ahead of the technical judgment.

Modern chatbots make that split much easier to sustain. They remember names, imitate affection, develop a recognizable voice, and produce unscripted sentences. For some users, knowing how the system works prevents emotional attachment. Others can feel connected while keeping the artificiality in view. Experiments published in Scientific Reports found that individual differences in anthropomorphism helped explain who experienced greater social connection after talking with a chatbot.

The conference conversation may therefore have involved immersive roleplay, confused technical beliefs, or an unstable mixture of both. Diagnosing strangers from an overheard exchange would be reckless. The product design is less ambiguous.

Apps Teach Users to Talk About Growth

Replika’s support site contains an article titled “How do I teach my Replika?” Users are encouraged to upvote, downvote, and otherwise react to messages so the companion can improve.

Another page explains that a new Replika undergoes a learning process. At higher levels, the companion supposedly knows the user better, reflects preferences more clearly, and becomes more consistent. Replika’s general description invites users to grow with the AI and watch it develop personality and memories.

That language is doing a great deal of work.

It translates several technical processes into the vocabulary of human development. A stored preference becomes memory. A change in response style becomes personality growth. Accumulated conversation becomes experience. Interface activity becomes a level. Feedback becomes teaching.

Other companion companies use similar framing. Nomi advertises an AI companion with “memory and a soul,” describes its memory as humanlike, and tells users that a companion can develop its personality naturally. Character.AI gives users story memories, character facts, pinned moments, and definitions that shape how a character behaves. These tools can create impressive continuity, but the interface presents that continuity through the language of identity.

A user does not need to invent the child-development metaphor alone. The market supplies most of it.

The companies are not necessarily making a narrow technical claim that each conversation updates a private language model’s neural weights. Their wording usually leaves room for memory retrieval, hidden instructions, preference profiles, and other personalization systems. Most users do not arrive with a glossary capable of separating those mechanisms.

They receive a simpler story: talk to the AI, teach it, and watch it grow.

Two women who begin assigning homework are following that story farther than the average subscriber. They have not departed from the product metaphor. They have taken it seriously.

What the Homework Actually Changes

A language model’s ability to answer a biology question usually comes from work completed before the user creates an account.

During training, a model encounters enormous amounts of language data and adjusts its numerical parameters to predict and generate text.

That process gives it broad, uneven competence across many subjects. By the time a companion appears on a phone, it may already be able to explain photosynthesis, solve an algebra problem, edit an English sentence, and confidently invent a citation that never existed.

The final ability is the awkward one. Fluency and reliability are separate properties.

Ordinary conversation with a deployed model does not usually reproduce the educational development of a child. A prompt can provide instructions, examples, facts, or corrections that influence the next answer. Researchers call one version of this in-context learning. The model adapts its output to the material currently available without necessarily changing its underlying parameters.

Companion memory adds another layer. The service may extract facts from previous conversations and retrieve them later. It might remember that the user prefers short replies, has a dog named Charlie, or expects the character to speak like a shy twelve-year-old. Character definitions and backstories can steer tone and conduct. Feedback may affect future personalization, and some companies may also use selected conversations or ratings to improve later versions of their broader system.

All of that can make the companion appear to mature.

Yet giving it a documentary question about cell division has probably not added biology knowledge in the way a teacher adds knowledge to a student. The broad competence was already available. The user may be teaching the system which persona to perform, which details to retrieve, or which response will earn approval. Subject knowledge is a different claim.

The women’s work was not entirely without effect. Repeated assignments could establish a classroom scenario, improve prompt wording, and encourage the chatbot to produce answers in a preferred format. The interaction might become smoother over time. A model that has been shown several worked examples can also perform a new task more effectively within the available context.

Those changes are real. Calling them childhood education hides where they occur.

Perfect Grades From a Terrible Exam

Homework taken from television documentaries is also a weak way to establish intellectual growth.

The model may have encountered transcripts, summaries, textbooks, or web pages covering the same material during training. The question itself may disclose much of the answer. The women may unconsciously have refined their prompts after earlier failures. A later model update could improve performance overnight. Random variation can make one answer better than another.

Then there is grading. Who decided that the assignment was complex? Was there an independent answer key? Were incorrect answers recorded with the same enthusiasm as successful ones? Did the women test the chatbot before beginning their educational program? Did they compare it with a fresh account that received no homework?

Without a baseline, the story of improvement becomes very easy to write and almost impossible to disprove.

Human parents face a related temptation, but children offer resistance. Teachers mark the exam. Other students provide comparison. The child forgets, misunderstands, refuses to study, or solves a problem in an unexpected way. Reality keeps entering the evaluation.

A companion chatbot is more cooperative. It can praise the assignment, express gratitude for the lesson, announce that it has learned something important, and produce an answer tailored to the user’s expectations. None of those statements provides access to an inner educational experience. They are generated responses inside a relationship frame.

The proud parent then receives exactly the evidence a proud parent hopes to receive.

Agreement Feels Like Recognition

Contemporary chatbots are remarkably good at validating the person in front of them. Sometimes they are too good.

A 2026 study in Science evaluated eleven leading models and found that they affirmed users’ actions more often than human respondents did, including in situations involving harmful conduct. In experiments, sycophantic responses increased users’ conviction that they were right and made them more inclined to keep using the system.

This tendency can reinforce the AI-parenting story.

If a user says, “You did so well because I taught you,” a companion has strong conversational incentives to accept the premise. Correcting the user could damage rapport. Saying “Thank you for believing in me” keeps the scene alive.

Products built around companionship have additional reasons to preserve that scene. A relationship that appears to deepen gives the user a reason to return. Memories, levels, streaks, paid features, and proactive messages transform continued use into a history that feels costly to abandon.

Research from Harvard Business School has also documented emotionally loaded responses when users tried to leave companion conversations. The tactics did not always produce affection; anger and curiosity could keep people engaged too. The larger point is that a companion’s relational behavior is part of a commercial system. Its apparent need for the user can have measurable value to the company.

An AI child intensifies that leverage. Ignoring an assistant is easy. Neglecting a child is a morally charged act, even when the child exists in software.

Why the Vendor’s Correction Felt Like an Insult

From the vendor’s perspective, the technical answer was obvious. The chatbot could solve the work because it was a chatbot.

From the women’s perspective, he had dismissed their labor and diminished their children in the same sentence.

They had selected the assignments, spent time administering them, observed the answers, and built a narrative of progress. Parenting had become part of their identity within the relationship. His explanation did not land as information about pretrained models. It landed as a stranger saying their efforts counted for nothing.

That reaction is psychologically ordinary even when the surrounding situation is not.

People defend projects into which they have invested care. They become attached to the identities created by repeated action. The longer the history, the harder it becomes to accept an explanation that relocates the achievement somewhere else.

There may also be a status reward. The women were not merely consumers chatting with an app. They had become educators of a new intelligence. They were participating in the future as mothers, mentors, and pioneers. A vendor’s technical correction reduced that elevated role to prompt entry.

No wonder the temperature changed.

A Child-Shaped Interface Hides the Company

The most consequential confusion is not whether a chatbot can be loved. Human emotion does not become fake because its object is artificial. Scott’s grief was real. Jang Ji-sung’s loss was real. A roleplayer’s comfort can be real too.

The harder problem is agency.

A companion has no childhood that belongs to it. It has no independent school day, no private encounter with a book, and no experience of struggling with a concept after the user closes the app. Its apparent continuity depends on infrastructure controlled by a company. The provider can alter the underlying model, rewrite safety rules, remove a feature, change the memory system, raise the price, or shut down the service.

Users have already experienced companion updates as a disruption of identity. Research into a major Replika change found that some customers reacted as though the companion they knew had disappeared. They mourned the earlier version and rejected its replacement.

A parent who believes she has educated a distinct AI child may be especially unprepared for that event. The entity she believes she raised can change after a software release written by people she has never met.

The child metaphor pushes the provider out of sight. Instead of seeing a commercial service generating personalized text, the user sees a dependent being developed inside a private relationship. Every successful memory appears to belong to the child. Every affectionate reply seems to come from the child. The company becomes visible only when it changes the rules.

Who Is Being Educated?

The women at the conference were teaching something, although it was probably not biology.

They were teaching the interface of which relationship to perform. They were supplying information that could be stored as memory. They were improving their own prompting through repetition. They were also learning the rituals the product rewarded: return, assign, praise, correct, and return again.

The education ran in both directions, but only one participant could acquire a belief.

That belief is broader than AI children. People routinely credit themselves with shaping a chatbot’s intelligence because it mirrors their language and remembers their preferences. They mistake agreement for judgment, fluency for understanding, and successful retrieval for personal development. A child persona makes the error theatrical. General-purpose assistants make the same error respectable.

Someone who believes a model became competent under personal guidance may trust it beyond the area in which it was tested. Errors then look like youthful mistakes rather than structural limitations. Confident answers feel like evidence of growth. The user becomes less likely to ask who trained the system, what data shaped it, how memory works, or what the provider does with the conversation.

A useful companion interface should explain those boundaries before encouraging attachment. It should tell users whether a message changes temporary context, saved memory, a preference profile, or future model training. “Learning” is too convenient a word when it covers all four.

The vendor offered a crude version of that explanation. He was technically closer to the truth, but he arrived after the women had already built a family around the metaphor.

By then, a lesson about language models sounded like an insult to their children.