In 1966, a computer science professor wrote a program of about two hundred lines. ELIZA matched patterns in a sentence and reflected them back as a question. It did almost nothing.
His secretary asked him to leave the room so she could talk to it privately. She knew exactly what it was. She was not deceived. She wanted the room anyway.
Sixty years of better pattern matching later, we keep asking the wrong question about that moment. The interesting fact was never that the program was persuasive. It was that the trade seemed worth making at all. Something was available in two hundred lines of pattern matching that was not available in the room she was already standing in.
The right question is not whether AI companions replace human relationships. It is which specific things people are taking from AI companions first, and what that ordering reveals.
Relationships are bundles
A friendship is not a single good. It is a package of at least eight separable components: attention (cognitive resources pointed at you), witness (someone holding a continuous record of who you have been), counsel (judgement applied to your situation), reciprocity (obligation running in both directions), risk (you can be hurt and so can they), physical co-presence (bodies in a room), social embeddedness (your friend knows your other friends and this constrains you both), and status conferral (being chosen by this person means something to others).
These components arrive bundled because, historically, they had to. There was no way to buy witness without also buying reciprocity. The bundle was a technological constraint dressed up as a metaphysical fact about what friendship is.
The unbundling
Bundles come apart when a technology makes it possible to sell one component separately. The newspaper was a bundle: reporting, classifieds, sports scores, obituaries, coupons. It arrived as one object. The internet unbundled it. Craigslist took the classifieds, which had been quietly subsidising the reporting for a century. What the unbundling revealed was not that people stopped wanting journalism. It revealed how much of the newspaper's economics had never been about journalism at all.
The AI companion is the first technology capable of selling attention, witness, and counsel detached from reciprocity, risk, embeddedness, and status. It cannot supply the second set. This is not a temporary engineering limitation. It is definitional. A machine cannot be hurt. Nothing was given up for you. The signal that makes care legible is missing.
What the ordering reveals
The strands being extracted first are precisely the strands that require time and sustained cognitive presence from another person. Nobody signs up to an AI companion for status conferral. Nobody is trying to acquire risk. The demand concentrates on attention, witness, and counsel.
This is a diagnostic reading, not a preference signal. When a substitute for part of a bundle is adopted rapidly and at scale, it identifies which part of the bundle was underserved, not which part was preferred. Rapid adoption of the first three relational strands, across people who already have intact social networks, reveals that those strands were underserved inside relationships that appeared, from the outside, to be complete.
A survey of student Replika users found the population lonelier than typical student samples while still reporting high perceived social support. These were not isolated people. They had people. The shortfall was specific: the attention economy has priced human attention out of human relationships. We built an industry that competes for the exact resource that friendship runs on, then expressed surprise when a technology that manufactures synthetic attention found a market.
Frequently asked questions
Can AI companions replace human relationships?
No, and that is not what they are doing. The more accurate description is relational unbundling: AI companions supply three of the eight components a friendship typically delivers (attention, witness, counsel) without supplying the others (reciprocity, risk, physical co-presence, social embeddedness, status conferral). The demand for the first three, detached from the rest, reveals that those were underserved inside existing relationships.
Why do people use AI companions when they already have friends?
A Replika user survey published in npj Mental Health Research found participants lonelier than typical student samples while still reporting high perceived social support. These were not isolated people. They had people. The shortfall was inside relationships that appeared complete from the outside, not in the number of relationships.
Is the AI companion industry exploiting loneliness?
Optimising for engagement on a product that enables attachment is a structural problem regardless of intent. Goodhart's law does the work: choose session length as the north star and the product will find dependency through a hundred small decisions no single person writes down. The question is whether any company in the category chooses a different metric.