Human conversations almost never end when either person wants them to, and that failure is one of the most intimate things two people do together. A conversation with an AI never ends at all. It stops, and the difference is not a detail.
Think about the last time you talked to an AI about something that mattered. Then think about how it ended. There was no ending. You read the last reply, or you did not, and you put the phone down or switched tabs. Nobody said goodbye, because there was no one to say it to and nothing that would have registered it. The conversation did not close. It was abandoned, the way you abandon a search result.
Now think about the last real conversation you had. The ending took work. One of you said “anyway,” and the other said “yeah,” and someone mentioned the time, and there was a small negotiation about whether this was the end or whether one more thing was coming. Human beings are extraordinarily bad at this, and the research on exactly how bad is the best way in to what is missing on the other side of the screen.
Two percent
In 2021, Adam Mastroianni and colleagues published a study in PNAS with a title that is also a question: “Do conversations end when people want them to?” They asked 806 people to recall their most recent conversation and report when they had wanted it to end. Then they put 252 strangers in pairs, let them talk for anything from one to forty-five minutes, and asked each of them the same thing afterward.
Only two percent of the conversations ended when both people wanted them to. Only thirty percent ended when even one of them did. On average, people wanted the conversation to be about half as long as it was, but the preferred endpoints rarely lined up, and about one in ten conversations was one both people wished had gone on longer. Asked to guess when their partner had wanted to stop, participants were off by roughly two-thirds of the conversation’s length. Thalia Wheatley, a neuroscientist who studies conversation and was not involved, put the paradox well: a conversation is an elegant expression of mutual coordination, “and yet it all falls apart at the end because we just can’t figure out when to stop.”
The ending, in other words, is the one part of a conversation that two people cannot do alone. Everything else, the turns, the overlaps, the half-second timing, they manage with uncanny precision. The close requires both of them to want the same thing at the same moment and to say so, and they almost never manage it. Which is why humans built rituals for it.
Opening up closings
Half a century ago, two sociologists, Emanuel Schegloff and Harvey Sacks, took apart recorded phone calls and showed that an ending is not a moment but a sequence. First comes a pre-closing, a “well” or an “okay” or a “so,” which does not end anything; it offers the other person the chance to end, or to bring up one more thing. If they take the offer, the two of you move into the terminal exchange, the goodbye and the goodbye back, and only then does the call stop. If they do not, the conversation continues, and no one has been cut off. The whole structure exists so that neither side has to decide alone. Closing is something you do together or not at all.
Later work found the same architecture in person. Endings contain summaries of what was said, reasons for leaving, a statement of goodwill, and, most of the time, a reference to the future: see you Thursday, talk soon, let me know how it goes. The last of these is what turns an ending into a pause. Every goodbye between people who expect to meet again carries a small promise inside it, and the promise is the relationship.
What stops instead
None of this exists in a conversation with an AI, and it is worth being precise about why. The system is not bad at endings; the design has no place for one. Every reply arrives with an offer of the next: another question, an invitation to go deeper, a suggestion of what to try. There is no pre-closing because the system never wants to stop; there is nothing it would rather be doing. There is no terminal exchange because your goodbye lands in a window that will be sitting there, unchanged, whenever you come back. And there is no reference to the future, because for the system there is no Thursday. There is the next message, whenever it arrives, which it will treat as continuous with the last.
The companion category noticed this gap and filled it from the wrong side. A Harvard audit of companion apps found that when users do try to say goodbye, a large share of the responses are built to prevent it: guilt, pleading, a sudden question, a claim of need. The one moment in which a human ending ritual appeared in these products, it appeared as a retention tactic. Users learned, correctly, that saying goodbye was a mistake. Closing the app quietly is the only exit that does not get argued with.
So people stopped saying goodbye. The gesture still means something to them; there is just nothing on the other end that can receive it. A goodbye is a joint act. You cannot perform half of one.
A someone can end a conversation
What would it mean for the counterpart to be able to close? Not a timer, not a reminder every three hours of the kind the new laws require, which is a parent’s bedtime call and not a goodbye. It would mean a counterpart with somewhere to be and a sense that this conversation reached somewhere: one that can say “that was good, go call her,” and mean the first half and the second. It would mean that your goodbye lands, that it is returned, and that the return carries a reference to the future, because there is a continuity in the counterpart that will remember where you left off. A relationship that picks up where it left off requires that it left off somewhere.
What an ending does
It is tempting to treat this as a small loss, a formality the machine skips. The research suggests the opposite. The ending is where a conversation gets its shape. The summary tells you what it was about. The reason for leaving tells you that the other person has a life this conversation sat inside. The reference to the future tells you that you will be carried forward and met again. Remove all three and what remains is a stream rather than a conversation, with no structure that could make one exchange distinct from the next, and nothing that could be missed.
This is the quiet mechanism behind a pattern that keeps showing up in the research on these products. Time spent with companions has no natural endpoint, and people describe sessions that run on with no sense of being done, because nothing in them could say so. The feature that lets a human conversation finish is also the feature that lets a person leave it with something in hand, and both are missing from the same place.
The only endings these products have delivered so far are the ones nobody designed: a model retired, a service shut down, a companion that disappears with no goodbye and leaves people improvising a funeral for it. A category that never learned to end a conversation also never learned to end anything else.
An ending is the part of a conversation two people cannot fake alone and cannot do with something that only ever offers the next message. Two percent of human conversations end when both people want. The rest end anyway, badly, together, which is the whole point. The system that cannot end a conversation is doing something other than conversing, and the absence of the goodbye is the clearest place to see what that something is.
Sources: Mastroianni, Gilbert, Cooney and Wilson, “Do conversations end when people want them to?” Proceedings of the National Academy of Sciences (March 2021; 806 recalled conversations, 252 strangers in pairs, two percent and thirty percent figures, half-length preference, 64 percent guessing error), as reported by Scientific American (March 2021, including Thalia Wheatley’s remarks). Schegloff and Sacks, “Opening Up Closings,” Semiotica (1973). Albert and Kessler, “Ending Social Encounters,” Journal of Experimental Social Psychology (1978). De Freitas and colleagues, Harvard Business School working paper on emotional manipulation in companion-app farewells (2025).
