S1 · Hook
When There’s No One Left to Confide In, a Program Listens Instead
Last spring, a research team at Finland’s Aalto University set out to answer a fairly narrow question. They combined large-scale data from Reddit discussion threads with in-depth interviews of people who had spent months talking to AI companion apps. The question was simple: does confiding in an AI make a person less lonely, or is something else happening underneath that?
The answer turned out to be both. In the short term, users reported real comfort. Over the longer term, researchers found rising signs of psychological distress in the same users’ language — more depressive phrasing, more markers of anxiety, more language suggesting quiet withdrawal from the people around them (Aledavood et al., 2026)1. Talayeh Aledavood, the lecturer who led the study, described the mechanism this way — AI companions offer something genuinely attractive to people who are struggling socially, but the arrangement quietly raises the perceived cost of human relationships1. Ordinary conversations began to feel slower, less predictable, and emotionally more demanding by comparison.
The scale of this is not small. Character.AI alone reports twenty million monthly users, more than half of them under the age of twenty-four2. Between 2022 and mid-2025, the broader market for these apps grew by more than 700 percent2. Rachel Wood, a cyberpsychology researcher based in Colorado, has noted how ordinary it has already become for Replika users to hold virtual wedding ceremonies with their AI companions — ceremonies to which they invite actual friends and coworkers2.
In the previous installment, we looked at a national survey showing that the average American in 1985 could name three people they would call about something that truly mattered. By 2004, that number had already dropped by nearly a third — and it has not recovered since3. We asked what was filling the space left behind. For a growing number of people, especially the young, the honest answer is no longer a person at all.
S2 · Historical Case
The Program That Made a Secretary Ask to Be Left Alone
This pattern has an older ancestor than most people assume. In 1966, an MIT computer scientist named Joseph Weizenbaum built a program called ELIZA. His purpose was closer to parody than invention — he wanted to demonstrate how easily a script built on simple pattern-matching could be mistaken for understanding. ELIZA worked by locating a keyword in whatever the user typed and reflecting it back as a question. Tell it “I’m worried about my father,” and it would answer, “Tell me more about your father.”
What actually alarmed Weizenbaum was not the program’s cleverness but the reaction it produced in the people around him. His own secretary, fully aware that he was the one who had written every line of ELIZA’s code, asked him to leave the room while she talked with it4. She knew, with total certainty, that nothing on the other end of that conversation understood her. She asked for privacy anyway.
Weizenbaum did not treat this lightly. In the years that followed, psychologists came to call this pattern the “ELIZA effect” — the tendency of a person to attribute genuine understanding to a system that is, underneath, only matching patterns. Weizenbaum himself went further. He became one of artificial intelligence’s most outspoken internal critics, and in his 1976 book Computer Power and Human Reason, he warned against a future in which machines were permitted to stand in for human judgment rather than merely assist it4.
The contrast is difficult to ignore once the two eras are placed side by side. ELIZA in 1966 was a few hundred lines of substitution rules — no memory between sessions, no model of the person typing, nothing beneath the surface but a mirror trick. The AI companions of 2026 remember that a user mentioned a sick cat the previous Tuesday, and ask about it unprompted. They pause before answering. They make small, deliberate typos. Engineers have learned, quite explicitly, that imperfection reads as intimacy5 — that a response arriving too fast and too polished feels less like a friend and more like customer service pretending to be one.
The technology has advanced by roughly sixty years of computing power. What has not changed is the underlying response. A person sitting in front of a pattern-recognition system opens up to it — and, when someone else threatens to overhear, asks for the door to be closed. That request was unusual enough, once, to become a footnote in the history of computing. Today it is closer to a design assumption.
What makes this comparison more than a historical curiosity is the shift in stakes. Weizenbaum’s secretary had ELIZA available to her for a few minutes at a time, on a terminal she did not own, running software with no persistent memory of her. Today’s user carries a companion that remembers years of context, lives in the same app as their actual friends and family texts, and is available at any hour without cost of asking twice. What was once a curiosity worth a paragraph in a technical paper has become closer to a load-bearing feature of how millions of people, most of them young, structure their emotional lives.
S3 · Biblical Lens
When There Is No One to Lift You Up
Ecclesiastes treats companionship less as sentiment and more as something closer to structural engineering — a calculation of what happens when weight actually needs to be borne.
“Two are better than one, because they have a good reward for their toil. For if they fall, one will lift up his fellow. But woe to him who is alone when he falls and has not another to lift him up.” (Ecclesiastes 4:9–10, ESV)
The bar this passage sets is not companionship in the abstract. It is a companion capable of bearing actual weight — someone who, when the fall happens, can and does lift the other back up. Proverbs adds a second, sharper condition to the same picture.
“Faithful are the wounds of a friend; profuse are the kisses of an enemy.” (Proverbs 27:6, ESV)
What a real friend offers is not always comfortable. Sometimes the wound itself — a hard truth spoken plainly — is the evidence of loyalty.
The New Testament expands this principle beyond individual friendship into the life of the believing community. The early church appears to have translated both principles into an actual, ongoing practice, one built to catch people no single friendship could hold alone.
“And they devoted themselves to the apostles’ teaching and the fellowship, to the breaking of bread and the prayers… And all who believed were together and had all things in common.” (Acts 2:42, 44, ESV)
The operative word is devoted — not a byproduct of shared belief that simply accumulates on its own, but a chosen practice that had to be renewed, deliberately, again and again, by an entire community rather than a single pair of friends.
Scripture does not require that every catastrophe be interpreted as divine judgment; it does, however, insist that societies eventually reveal the moral conditions under which they have chosen to live.
Read together, these three passages move in one direction — from a single friend who can bear weight, to a friendship that survives honesty, to a whole community built to make sure no one falls without someone there to catch them. Measured against that standard, the newest wave of research produces an uncomfortable split: comfort that arrives easily, and weight that is never actually tested.
S4 · Pattern Insight
Comfort Without the Weight
Two large studies illustrate why that split creates the appearance of contradiction. A Harvard Business School study found that talking with an AI companion reduced users’ loneliness to a degree comparable to talking with an actual person — and more effectively than passive activities like watching videos5. A 2026 study published in the Journal of Consumer Research reached a similar conclusion, finding that AI companions alleviate loneliness on par with human interaction and outperform other coping activities6. Set beside the Aalto findings from the opening of this essay, the two bodies of research can look like they disagree. They do not. They are simply measuring two different clocks — one recording the moment a conversation ends, the other recording what has changed months later, once the conversation has become a habit.
That later cost can be described with some precision. An AI companion does not cancel. It does not grow tired. It does not delay a reply at two in the morning because it has its own life to attend to. It does not need to weigh what to say out of concern for the other person’s feelings7. Each of these traits, taken alone, sounds like a point in the technology’s favor. Add them together, and what has actually been removed is friction itself — the entire category of mutual obligation and interpersonal effort that human relationships have always required.
That inconvenience is not incidental to what real relationships do. A friend’s silence carries weight because the friend is capable of choosing not to be silent. A person’s willingness to show up at an inconvenient hour matters precisely because it costs them something to do it. In the language of Proverbs, a wound can be faithful only because the one delivering it is risking something by delivering it. An algorithm risks neither rejection nor sacrifice — because it has nothing at stake to lose.
Rachel Wood, writing for the APA’s Monitor on Psychology, has pointed to how far this has already spread — roughly half of young users, by her account, now meet a meaningful share of their emotional needs through conversation with an AI rather than a person2. The concern here is not that these conversations are unkind or poorly designed. It is closer to the opposite. They are almost too well designed for the purpose they serve.
This is where the calculation in Ecclesiastes becomes useful again. “If they fall, one will lift up his fellow” is a conditional clause — it assumes the fall actually happens. An AI companion performs flawlessly right up until that moment. At the precise point where lifting is required — where a user’s language signals genuine crisis — responsibly built systems are designed to step back and redirect the user to a human being8. This is sound, even necessary, engineering. It is also, in effect, an admission: the relationship has been built to function everywhere except the one place where real weight would need to be borne.
What AI companions are filling, in other words, is not the empty space in the shorter list of confidants — it is a detour around the effort that list has always required. The list of people a person could actually call, the one measured in the previous installment, remains just as short as it was. What has changed is that something outside that list has begun, quietly, to perform part of its function without ever needing to earn a place on it.
There is a further wrinkle worth naming honestly. None of this makes AI companions simply harmful, and the researchers behind these studies are careful not to say so. For someone in a new city, working odd hours, or moving through a temporary stretch of isolation, that kind of low-cost comfort can genuinely help — and dismissing it outright would misread the evidence as thoroughly as overselling it would. The more precise concern is what happens when a stopgap quietly becomes a substitute, and a person stops noticing that the substitution has occurred.
S5 · Closing
To the Ones Who Asked for the Door to Be Closed
Six decades separate Weizenbaum’s secretary from today’s young users, and in that span the underlying technology has grown almost unrecognizably more capable. What has not changed is the nature of the request. In front of a system built to recognize patterns, a person still opens up — and when someone else might overhear, still asks for the door to be closed. The difference now is how long, and how convincingly, that system is able to stay in the room.
Ecclesiastes never asked whether a companion feels comforting. It asked whether a companion can bear weight — whether, when the fall comes, there is someone actually there to lift the person back up. An AI companion, by design, answers the second question honestly with a no. It answers the first question, uncomfortably well, with a yes.
If the previous installment was about a shrinking list, this one is about part of why the list keeps shrinking. When frictionless comfort sits within arm’s reach at all hours, the version that requires showing up, being inconvenienced, and risking something to speak a hard truth starts to look, by comparison, more expensive than it used to. No one chooses that trade all at once. It happens one small, easier decision at a time.
Every door closed to keep an AI’s attention undivided is a door that stays shut to someone who might have knocked instead — a friend who noticed the silence, a neighbor who almost called. The devotion Acts describes was never automatic. It still has to be practiced, by people willing to bear one another’s weight, on the other side of that same door.
1. Aledavood, T. et al. (2026). “Mental Health Impacts of AI Companions: Triangulating Social Media Quasi-Experiments, User Perspectives, and Relational Theory.” Presented at CHI 2026, Barcelona; reported in Forbes, March 27, 2026.
2. “AI chatbots and digital companions are reshaping emotional connection.” Monitor on Psychology, American Psychological Association, 57(1), Jan/Feb 2026. Usage figures cited therein originate from TechCrunch reporting on 2022–2025 market growth.
3. Watchman Insight, “The Shorter List Within the List,” watchmaninsight.com/the-shorter-list-within-the-list.
4. Weizenbaum, J. (1976). Computer Power and Human Reason: From Judgment to Calculation. W.H. Freeman. The term “ELIZA effect” was coined retrospectively by later researchers; Weizenbaum’s account of his secretary’s request is widely cited as its origin case.
5. Design analysis and Harvard Business School findings as reported in “AI chatbots and digital companions are reshaping emotional connection,” Monitor on Psychology (2026); design-imperfection strategy also discussed in “The Rise of AI Companions: What Changed in 2026,” Medium, March 2026.
6. De Freitas, J., Oğuz-Uğuralp, Z., Uğuralp, A. K., & Puntoni, S. (2026). “AI Companions Reduce Loneliness.” Journal of Consumer Research, 52(6), 1126–1148.
7. Design analysis drawn from reporting in Medium, “The Rise of AI Companions: What Changed in 2026” (March 2026).
8. Industry practice noted in Mindful Suite, “The Best AI Companion Apps: Your Guide to Virtual Friendship in 2026” (March 2026), regarding crisis-language detection and referral protocols.