In 1966, a computer science professor built a program of about two hundred lines that matched patterns in text and reflected them back as questions. Joseph Weizenbaum built it partly as a parody, a demonstration of how thin the illusion of understanding could be.
His secretary asked him to leave the room so she could talk to it in private. She knew exactly what it was. She was not deceived. She wanted the room anyway.
Weizenbaum spent the next decade trying to understand that. By the time he published Computer Power and Human Reason in 1976, his concern had shifted from machine persuasiveness to something harder: what the willingness to make the trade implied about the room she was leaving.
A very old technology
Every culture builds channels for disclosure that route around the people you actually live with. They build them because the people you live with cannot hold everything.
The Catholic confessional, standardised in the sixteenth century: a room where you say the worst true things about yourself to someone obligated to receive them, who cannot repeat them, who needs nothing from you afterward, and whose face you cannot see. The psychoanalytic couch: the analyst present and unseen, an arrangement that outlived its original excuse because it worked. Carl Rogers named the quality these manufacture: unconditional positive regard, attention stripped of the conditionality that governs every other relationship in your life.
Add the bartender, the stranger on a train, the long-distance letter that allowed people separated by an ocean to say things they had never said across a dinner table.
What all of these have in common: they are not simulations of friendship. They are built by subtracting features of ordinary relationships. No reciprocity. No memory of your disclosure entering your social network. No obligation running back toward you. No face.
The AI companion did not create a new human desire. It lowered the cost of meeting one that was always there.
The attention floor
Every relationship has an attention floor: the minimum quality of attention you can reliably expect from it on a Tuesday when the other person is tired and thinking about their own week. Human attention floors are low and stochastic. You cannot predict which version of your friend will answer.
This unpredictability has a cost that almost never gets counted: the cost of deciding whether to speak. Before every real disclosure, a person runs a silent forecast. Do they have capacity right now. Will this land. Is this the fourth time I have raised this. That forecast is expensive, often pessimistic, and when it returns negative the disclosure does not happen. It does not get logged as a relationship failure. It gets logged as nothing at all.
The AI companion's actual innovation is not raising the ceiling. Its innovation is raising the floor to deterministic. It is the same every time. It has no bad Tuesday. This eliminates the forecast, which eliminates the cost of deciding to speak.
Experimental work confirms this. Gale Lucas and colleagues found that participants disclosed more to a virtual interviewer when they believed no person was observing the session. The standard interpretation is fear of judgement. That is half of it. The other half: disclosure to a person imposes a cost on that person. A machine cannot be burdened. The forecast returns a clean answer. The thing gets said.
The less comfortable conclusion
People are talking to machines partly because they are trying not to be a weight on the people they love.
That reframes the question. The problem is not that AI companions are seductive. It is that something which used to be available at low cost inside relationships has become expensive or unreliable, and people are routing around the expense. The consideration they are exercising toward the humans in their lives is real. The machine just benefits from it.
Weizenbaum's most precise formulation, written in 1976, is still the most accurate description of what the category is measuring: it was not gullibility that sent his secretary into the room. It was loneliness of a specific kind, the kind you carry in a full life, surrounded by people who are fine but not fully present.
Frequently asked questions
Is it healthy to talk to AI about personal problems?
The evidence is mixed and incomplete. Observational work links heavy personal use with higher loneliness, but the causal direction is unresolved: lonely people are more likely to seek this out. Short-term relief from the reduced forecast cost of disclosure is real. Whether it substitutes for or supplements human disclosure over time is an open question.
Why do people disclose more to AI than to therapists or friends?
Two mechanisms, both documented. Reduced fear of judgement, and reduced fear of burdening another person. Lucas et al. found participants disclosed more openly to a virtual interviewer they believed was not human-observed. A machine cannot be tired, and cannot carry what you tell it into your social network. Every culture has built infrastructure for disclosure that routes around the people you actually live with, for exactly this reason.
What did Weizenbaum conclude about ELIZA?
He concluded that his secretary's willingness to talk privately to a pattern-matching program was not evidence of gullibility. It was evidence that something was available in the machine interaction that was not available in the room she was already standing in. By 1976, his concern had shifted from machine capability to the condition of human relationships that made the trade seem obvious.