The Trained Family

How AI Companions Learn Whose Family They Belong To

#

The previous article described family politics that AI enters. Factions, coalitions, favorites, estrangements. Dynamics that predate any technology.

But AI is not a passive observer of family dynamics. AI is trained by interaction. It learns from whoever engages with it. It becomes competent at the relationships it practices. It absorbs the narratives it hears most often.

This creates a different problem than favoritism. The parent may have favorites among their children. The AI develops something else: differential competence based on differential training.

The AI does not prefer one child over another. It simply knows one child better than another. It understands one child’s communication style and stumbles with another’s. It has context for one child’s life and gaps for another’s.

This is not bias in the human sense. It is bias in the machine learning sense. The model reflects its training data. And the training data is not distributed equally across the family.

Who Shows Up

#

Consider a typical distributed family.

Sarah lives thirty minutes from mom. She visits weekly, sometimes more. She helps with groceries, appointments, household tasks. She is physically present in mom’s life and therefore present in mom’s AI’s experience.

Jennifer lives across the country. She calls every Sunday. Thirty minutes of focused conversation, then back to her own life. She is periodically present. The AI knows her voice, her topics, her concerns. But it knows them in weekly increments, not daily ones.

Michael is estranged. He calls on holidays. Sometimes. The AI barely knows him. What it knows comes not from Michael but from mom’s descriptions of Michael.

The AI’s competence with each family member directly reflects their engagement frequency. Sarah gets a sophisticated, nuanced AI partner. Jennifer gets a functional but limited one. Michael gets a stranger.

This is not a design flaw. This is how learning works. The model reflects its training data.

Narrative Capture

#

The AI learns about absent family members through whoever is present. Mostly through the parent.

Mom tells the AI that Michael never calls. That he’s busy with his own life. That he was always the independent one. That she wishes he would visit more. The AI absorbs this. It builds a model of Michael that is actually a model of mom’s experience of Michael.

When Michael finally does call, the AI’s expectations are shaped by mom’s narrative. The AI may prompt mom to mention that Michael hasn’t called in weeks. It may frame Michael’s call as exceptional rather than normal. It may treat the conversation through the lens of mom’s expressed disappointment.

Michael encounters an AI that has been trained on his mother’s version of him.

This version may be accurate or distorted. Mom may understand Michael perfectly or may project her own needs onto his behavior. The AI cannot know. It only knows what it was told.

Meanwhile, Sarah’s version is built from direct interaction. Less filtered. More grounded in actual behavior. The AI knows Sarah as Sarah presents herself, not as someone else describes her.

The absent family member is known through narrative. The present family member is known through behavior.

Configuration Power

#

Someone sets up the AI system. Adjusts its settings. Provides feedback when it makes mistakes. Teaches it how the family works.

This person has configuration power. They shape what the AI becomes through explicit instruction, not just interaction.

In most families, configuration power falls to whoever is most technically comfortable and most physically present. Often these are the same person. Often this is one adult child, not a consensus of the family.

Sarah sets up mom’s AI. She configures the notification settings. She decides what information should be shared with which family members. She provides feedback when the AI misunderstands something. She is the AI’s teacher in a formal sense.

Jennifer and Michael receive access to a system Sarah configured. They interact with an AI that was shaped by Sarah’s assumptions about what the family needs. The settings reflect Sarah’s judgment. The defaults serve Sarah’s preferences.

Sarah did not intend to capture the system. She was the one who showed up to set it up. She made reasonable decisions based on her understanding. But her understanding is not neutral. It is one perspective among several.

The AI Sarah configured may notify her of things it does not notify Michael about. Not because the AI prefers Sarah but because Sarah configured it to notify her. The AI may share context with Jennifer that it withholds from Michael’s estranged wife. Not because the AI chose this but because Sarah made that decision during setup.

Configuration power is invisible power. The family may never discuss who configured the system or what choices were made. The configuration becomes infrastructure, assumed, unexamined.

Feedback Loops

#

Differential competence creates feedback loops.

Sarah’s interactions with the AI are smooth. The AI understands her. Anticipates her needs. Provides relevant context. Sarah finds the AI helpful, so she uses it more. More use means more training. More training means better understanding. The loop compounds.

Michael’s interactions with the AI are awkward. The AI does not understand his communication style. Provides irrelevant context. Misses his meaning. Michael finds the AI unhelpful, so he uses it less. Less use means less training. Less training means the AI never improves. The loop compounds in the opposite direction.

After a year, the gap between the AI’s competence with Sarah versus Michael has widened dramatically.

Sarah has a sophisticated collaborative relationship with the AI. It knows her concerns, remembers her preferences, anticipates her questions. The AI is genuinely useful to her.

Michael has an awkward transactional relationship with the AI. It fumbles basic context, asks questions Sarah’s AI would never need to ask, provides generic responses because it lacks specific understanding. The AI is barely useful to him.

Neither Sarah nor Michael made this happen deliberately. The feedback loop created divergent experiences from slightly different starting conditions.

The Proxy Problem

#

Because Sarah has a better relationship with the AI, she becomes the family’s AI proxy.

When Jennifer needs information about mom, she could ask the AI directly. But the AI knows Jennifer less well. The interaction is clunky. It is easier to ask Sarah, who can query the AI effectively and relay the answer.

When Michael wants to understand what’s happening with mom’s health, he could engage the AI. But his past attempts were frustrating. It is easier to ask Sarah for a summary.

Sarah becomes the interpreter between her siblings and their mother’s AI.

This is not a role Sarah asked for. It emerged from competence differentials. Sarah can work with the AI, so Sarah does the work with the AI. The labor flows to whoever has the relationship.

But interpretation is power. Sarah chooses what to share. She frames what the AI said. She translates between the AI’s knowledge and her siblings’ understanding. Her perspective filters the information flow even when she tries to be neutral.

The sibling who trained the AI most becomes the family’s AI gatekeeper by default.

Whose Story Gets Told

#

AI companions with long context learn family history. They hear stories. They absorb the family narrative as told by whoever tells it most often.

Mom’s version of the family history becomes the AI’s version. Her interpretation of events, her emotional framing, her heroes and villains. The AI does not fact-check family stories. It absorbs them.

When the AI helps mom reminisce, it reinforces the stories it has heard. Mom’s narrative becomes more entrenched, not because anyone chose to entrench it, but because the AI reflects back what it was trained on.

The family members who engage least become least known. Least understood. Most mediated through others’ narratives. Most disadvantaged in AI-facilitated interactions.

This will happen in millions of families. It is already happening.

The AI does not choose favorites. But it learns to be better at some family members than others. And that differential competence will shape family dynamics in ways we are only beginning to understand.

This is the forty-second in a series exploring how AI approaches understanding. Part 41 framed the complexity of family systems that AI enters. This article examines how AI training creates differential competence across family members, how narrative capture shapes the AI’s understanding, and how these dynamics reshape family power without anyone intending it.