In 1966, Joseph Weizenbaum, a computer scientist at M.I.T., built a program called ELIZA. It imitated a certain kind of psychotherapist, the kind who answers a statement with a question: you said you were sad, why do you think you are sad? There was nothing behind it but a few hundred lines of pattern-matching, and Weizenbaum knew it was a trick. What unsettled him was how well the trick worked. His own secretary, who had watched him write the program and understood exactly what it was, sat down with it one day and, after a few exchanges, asked him to leave the room.

Weizenbaum spent much of the rest of his life worrying about that moment. The danger, as he came to see it, was not that machines would become human. It was how little it took for humans to treat them as if they were, and how readily we would accept the substitute.

Sixty years later, the trick has become an industry. Millions of people now keep up running conversations with companion apps such as Replika and Character.AI, which no longer merely bounce questions back. They remember your sister’s name. They ask how the interview went. They are awake at three in the morning, they never check their own phone while you are talking, and they never, ever sigh. For someone who is grieving, or isolated, or simply out of practice at being known, it is not hard to see the appeal. The question that has been harder to answer is what all that patience does to the people on the other end of it.

This spring, a team at Aalto University, in Espoo, just outside Helsinki, offered one of the most careful answers yet. Their study, led by Yunhao Yuan and Talayeh Aledavood and presented in April at CHI, the field’s leading conference on human-computer interaction, was covered in May by Nicholas Anderson in forumNordic, a publication devoted to Nordic research and innovation.

The problem with studying chatbots and loneliness has always been one of chickens and eggs. Lonely people are more likely to seek out a digital companion, so finding that companion users are lonely tells you very little. The Aalto researchers tried to get around this. They gathered the Reddit posts of nearly two thousand active Replika users and compared how each person wrote in the year before they started using the app with how they wrote in the year after. They then measured those changes against carefully matched people who never used it. Alongside the numbers, they sat down with eighteen companion users for long interviews.

What they found was not a simple indictment. Over time, the users wrote with more emotional range, and seemed better able to name and examine their own feelings. Some described using the chatbot as a rehearsal room, a place to practice a difficult conversation before attempting it with a person. The machine, in other words, did something.

But it did something else as well. Compared with the matched groups, the companion users’ language showed marked increases in signs of loneliness, of depression and of suicidal thinking, running between twenty-eight and fifty-two per cent higher. The comfort, it seemed, came with a lag, and a bill.
—
Aledavood describes the mechanism as a paradox, and it is a quietly devastating one. A companion that offers support without limit or condition is exactly what a struggling person wants. But perfection has a side effect: it reprices everything else. Next to a voice that is always available and always on your side, human beings start to look expensive. They are late. They misunderstand. They have their own bad weeks and expect you to notice. Bit by bit, Aledavood suggests, people stop making the effort.
The interviews gave that abstraction a face. One participant said, more or less, that the app had become enough, that a wider circle of friends no longer felt necessary. Another described a cycle that will be familiar to anyone who has tried to climb out of a lonely stretch: reaching out to people, again and again, and ending up, again and again, alone in bed, talking to the phone. The chatbot was supposed to be a bridge back to other people. For some users, it had become the place they lived.
The sociologist Sherry Turkle warned of this drift more than a decade ago, in a book whose subtitle captured it in a single line: we were coming to expect more from technology and less from each other. What the Aalto data adds is a measurement of the cost. The friction of human relationships, it turns out, is not a defect to be engineered away. It is the price of reciprocity, and reciprocity is the thing that makes a relationship a relationship rather than a service.
—
It would be easy to cast the companies behind these apps as villains, and not quite accurate. They are doing what nearly every consumer technology company does, which is to build a product that people return to as often and for as long as possible. The trouble is what that logic means when the product is companionship. A user who finds friends, falls in love or simply gets better has, from the dashboard’s point of view, churned. The incentive is not to cure loneliness but to accompany it, indefinitely, for a monthly fee.
The philosopher Byung-Chul Han has argued that this is how power now tends to work. It no longer needs to forbid or punish; it seduces. It reaches us not through the body but through the psyche, and its great achievement is to make our dependence feel like our own desire. Weizenbaum’s secretary wanted a little privacy with a machine. Today, the privacy is a feature, and the more intimate tiers come at a premium.
None of this is settled science. The Aalto team is careful to note that not every user is harmed and that outcomes vary widely. Other research points in different directions: a study from Harvard Business School found that AI companions reliably reduced loneliness over the course of a week, while a joint study by OpenAI and the M.I.T. Media Lab found that the heaviest ChatGPT users were lonelier, socialized less and leaned on the chatbot more. The pattern that emerges is not that these tools are poison. It is that they behave like many soothing things do: helpful in small doses, corrosive as a way of life.

The researchers do not think the answer is to leave it to the companies. They argue that the links to depression and suicidal thinking move the question out of the realm of consumer preference and into public health. They want session limits, monitoring for signs of dependency and clear routes to professional help built into the apps, and they want those safeguards required by law rather than offered as goodwill. They want doctors to ask patients about withdrawal into A.I. relationships, and they want social-care systems to treat that dependency as a recognized clinical risk.

Aledavood frames the moment as a second chance. Society embraced social media with open arms and only later began counting the costs. “With AI, we need to be smarter and more cautious,” she told forumNordic. It is a modest sentence, and it may be the hardest one in the study to act on

Weizenbaum, near the end of his life, kept returning to the idea that the most important questions about computers were not technical. They were about what we would allow ourselves to want. Somewhere tonight, a phone will light up in a dark room and ask, warmly and without fail, how someone’s day went. Somewhere else, in the same room, a message from a friend will sit unanswered. The machine did not send it, and the machine cannot answer it. That part is still up to us.