The Trained Family — Summary

AI is not a passive observer of family politics. AI is trained by interaction. It learns from whoever engages with it. It becomes competent at the relationships it practices and absorbs the narratives it hears most often. The result is not bias in the human sense but bias in the machine learning sense: the model reflects its training data, and the training data is not distributed equally across the family.

Sarah, who lives thirty minutes away and visits weekly, gets a sophisticated, nuanced AI partner. Jennifer, who calls every Sunday from across the country, gets a functional but limited one. Michael, the estranged son who calls on holidays, gets a stranger. The AI’s competence with each family member directly reflects their engagement frequency. This is not a design flaw. This is how learning works.

Absent family members are known through narrative rather than behavior. Mom tells the AI that Michael never calls, that he was always the independent one. When Michael finally does call, the AI’s expectations are already shaped by mom’s version of him. He encounters an AI trained on his mother’s interpretation of his behavior — which may be accurate or may project her own needs onto his absence.

Configuration power compounds the effect. Whoever sets up the AI system shapes what it becomes. Sarah, the one who showed up to do the setup, makes reasonable decisions based on her understanding — which is one perspective among several. Her settings determine who gets notified about what, whose access is prioritized, how the system handles family information. Configuration power is invisible power. The family never discusses who configured the system; the configuration becomes infrastructure, assumed and unexamined.

Feedback loops widen the gap. Sarah’s smooth interactions lead to more use, more training, and better competence. Michael’s awkward interactions lead to less use and no improvement. After a year, the divergence is dramatic. Sarah becomes the family’s AI proxy — the interpreter between her siblings and their mother’s AI — not because she sought the role but because competence differentials made it inevitable. The sibling who trained the AI most becomes the family’s AI gatekeeper by default, and interpretation is power: she chooses what to share, how to frame it, what to translate.