Quick takeaways
|
Every evening hands you a small, repeated choice. Text a friend who might not respond for hours, who might be busy, who might want something complicated and tiring of you in return. Or open a chat window that responds instantly, asks nothing back, and never makes you feel like a burden. Multiply that choice across a year, and you get an answer to a question nobody consciously decided to ask: where did all your time and attention for other people actually go.
This is not really a story about AI replacing human relationships in one move. It is a story about what happens when an easier option sits next to a harder, more valuable one, over and over, until the harder option quietly stops getting chosen. The broader research connecting heavy AI chatbot use to worsening loneliness over time points at the same underlying mechanism this article is about to walk through in detail.
The substitution hypothesis researchers keep finding evidence for
There is a name for this in the research literature: the social substitution hypothesis. It holds that reliance on AI companions does not simply add a new form of connection alongside existing human relationships. Instead, it displaces opportunities for the higher quality human contact that would otherwise have happened. This sits in deliberate contrast to a more optimistic idea called the social compensation hypothesis, which predicts that people with weaker social networks would use chatbots to fill a real gap and come out ahead.
Large scale research testing both ideas against each other found a result that favored substitution over compensation. People with smaller social networks were indeed more likely to seek companionship from chatbots, but using chatbots for companionship did not improve their well-being relative to people who avoided that kind of use. The researchers behind that study found a negligible interaction between social network size and chatbot companionship, meaning the chatbot was not meaningfully closing the gap for people who needed it most. Companionship-oriented use was, on the whole, associated with lower well-being across the board.
📋 Quick note Compensation and substitution are competing predictions, not two ways of describing the same thing. Compensation says the chatbot fills a real gap. Substitution says the chatbot eats the time and motivation that would have gone toward closing that gap with an actual person. The evidence currently leans toward substitution. |
Why the easier option always has an advantage
A human relationship requires reciprocal effort to stay alive. You owe your friends responses, attention, the occasional inconvenience of showing up when you would rather not. That obligation is not a flaw in human relationships. It is the mechanism that makes them valuable, the thing that signals you actually matter to someone enough for them to put something in.
An academic Voices article on this exact topic identified reduced human-human socialization as one of several plausible downstream harms of chatbot use, specifically because conversational AI can provide relational benefits without requiring users to make the reciprocal effort a human relationship demands. That asymmetry is the entire mechanism in miniature. Given a choice between an option that costs nothing and one that costs real effort, people will reliably choose the free option more often, even when the costly option pays off better in the long run. Nothing about that requires AI to be malicious or even particularly good. It only requires AI to be available and frictionless.
The cost difference that drives the choice | |||||
|
| ||||
The part nobody expected: it can change how you see other people
The most genuinely surprising finding in this research area is not about time at all. It is about perception. A pair of preregistered studies found that people who had a brief interaction with an AI, compared to a purported human, became more demanding and instrumental in how they treated their conversation partner, and showed less positive affect while doing it. The unexpected part came next: that more demanding, less generous attitude spilled over into a completely unrelated task, coloring how those same participants judged another, actual human being immediately afterward.
The interactions in question lasted under three minutes and had nothing to do with the later evaluation. The researchers concluded that interacting with AI may be quietly training people toward a more transactional way of relating, and that habit does not stay neatly contained to the AI conversation. It leaks into how people treat each other.
How a 3 minute AI chat changed behavior | |
| 01 | Participants had a brief, unrelated interaction believing their partner was AI. |
| 02 | They became more demanding and instrumental, with less positive affect, during that exchange. |
| 03 | That harsher attitude spilled into how they judged an unrelated human afterward. |
Spend enough time in a relationship that asks nothing of you, and ordinary human relationships, which by their nature ask quite a lot, can start to look unreasonably demanding by comparison. The chatbot is not just taking up time that could have gone to a person. It may be quietly recalibrating what feels like a normal amount of effort to expect from one.
⚠️ Common mistake Assuming the crowding out only happens through lost time. The comparison effect on perception, where people start to look less appealing relative to a frictionless chatbot, is a separate mechanism that can operate even when total hours spent with people have not changed much yet. |
This is not a new problem, just a faster moving version of one
This displacement logic is not unique to AI. Researchers studying the early days of internet adoption were already asking whether time online displaced time spent socializing in person, and found, in a sample of more than six thousand Americans, that time spent using the internet at home was negatively associated with time spent socializing with friends. AI chatbots did not invent the substitution mechanism. They arrived as a faster, more emotionally convincing version of a pattern researchers had already been tracking since the late 1990s.
What changes with AI is the quality of the substitute. Early internet use displaced in-person time with something thinner, browsing, forums, asynchronous messages. A chatbot displaces that same time with something that can simulate warmth, memory, and personalized attention convincingly enough to feel like it is meeting the need it is actually just absorbing. The same dynamics reshaping how much Americans trust AI systems more broadly are at play here in a more intimate register: a system convincing enough to feel trustworthy can end up substituting for judgment, attention, or company that used to require another person.
Common misconceptions about this crowding out effect
Myth: this only happens to people who use AI companion apps obsessively. The displacement mechanism operates gradually through ordinary, moderate use. It does not require obsession, only a consistent pattern of choosing the lower friction option often enough that it adds up.
Myth: more time with a chatbot automatically means less time with people. The relationship is not perfectly one for one, and some chatbot use genuinely fills idle time that would not have gone to socializing anyway. The research concern is specifically about companionship-oriented use crowding out higher quality contact, not all AI use displacing all human contact equally.
Myth: the comparison effect on judgment is a minor side note. It is one of the more striking findings in this area precisely because it is invisible to the person experiencing it. Nobody notices themselves becoming slightly more demanding of others; they just notice that people seem more tiring than they used to.
Myth: this is fundamentally different from older technology displacement effects. The mechanism researchers identified in early internet studies, time spent on a technology displacing time spent with people, is the same basic pattern. What is different is how convincing the substitute has become.
What the crowding out actually costs
People rarely wake up and decide to replace their friendships with a chatbot. The substitution happens through a long series of small, individually reasonable choices, the easier option taken often enough that it quietly becomes the default, while the harder option simply gets chosen less and less until it barely gets chosen at all. The relationships that needed reciprocal effort to survive do not get a dramatic ending. They get neglected by inches.
Recognizing the mechanism does not require giving up AI as a tool. It requires noticing the specific moments where the easy option and the valuable option are competing for the same slot in your evening, and treating that competition as a real cost rather than a free choice. The research on early internet displacement suggests this pattern was never really about the technology itself. It is about what people do when something effortless is placed next to something that was always supposed to take effort, and the academic case for studying these downstream harms more closely makes clear that researchers think this question deserves a lot more attention than it has gotten so far.
