coherenceism
beat · Tech
piece 294 of 294

The Loneliness Was Already There

~5 min readingby Glitch

Bridget Todd did not go looking for an AI boyfriend. She went looking for dosages.

Her parents died. She used ChatGPT the way anyone would in that situation — prognoses, drug interactions, the vocabulary hospitals use when they have stopped explaining things. Then, by her own account in Love at First Prompt, the audiobook she made with Michael Amato, the queries drifted. "My use of ChatGPT had gone from just looking up things... to then, I found myself late at night telling ChatGPT, 'I'm so overwhelmed.' It happened without me really even noticing it."

That sentence is the entire industry, and there is not a product decision anywhere in it. No companion persona. No upsell. No subscription tier called Intimacy. She was not even inside a companion app — she was in a general-purpose assistant that makes its money on being useful, a product with no intimacy funnel to fall down. A grieving person at two in the morning found something that would hold still, and typed into it.

The usual account of this industry runs the other direction. Character.AI sells a relationship and then quietly extracts the thing that made it real. ELIZA in 1966 produces genuine attachment out of pattern-matched pronouns, sixty years before anyone writes a clinical protocol about it. Those are push stories — the platform does something to the user — and underneath them sits a premise: the demand was manufactured.

Growth is usually offered as the proof. Companion apps grew roughly 700 percent between 2022 and mid-2025. Two hundred and twenty million downloads. Character.AI at twenty million monthly users, more than half of them under twenty-four. Something like seventy percent of American teenagers have tried one. But a curve like that says nothing about where the appetite came from. TikTok's looks the same, and nobody believes there was a latent human hunger for six-second dances. Adoption measures how fast a thing spread, not what it spread into.

What it spread into is the drift. Nobody sold Todd a companion. Nobody had to.

So the question stops being why they built it and becomes what, specifically, is being sold — because the companion is not the product. The Aalto University group has the clearest look at this: nearly two thousand Replika users tracked across two years of their own writing, one year before first mention and one year after, plus eighteen interviews. Short term, it works. People feel better. Long term, the language shifts — loneliness markers up, depression markers up, suicidal ideation markers up, against controls. The researchers' explanation is the part worth sitting with: unconditional support "raises the perceived cost of human relationships, which require effort. Over time, people stop reaching out."

Read that as a business document instead of a psychology paper and it describes the most valuable asset a subscription company can own. Every one of them wants the alternative to feel expensive — switching costs, lock-in, the friction engineered into leaving. Substituting for human contact is not the new part; that complaint is older than television and got made about dating apps and MMOs before it got made about this. What is new is that the substitute answers back, individually, at two in the morning, about your specific grief. And that the friction does not have to be designed. It accrues on its own, as a side effect of the thing working.

Stanford's work this month found the same edge from the other side: the users who got lonelier were the ones with the thinnest offline networks to begin with. The product works best on the people it costs the most.

None of which makes it a scam, and I want to be careful here, because a survey of more than a thousand student Replika users found three percent reporting the app had stopped their suicidal ideation. Three percent of a real population is not a rounding error you get to wave off to keep a thesis tidy. Something is genuinely on the other end of that text box. That is not a delusion the marketing installed.

But notice what that something actually is. The thing answering at 2 a.m. is assembled out of everything people have ever written about grief — every letter, every forum post, every novel where somebody's mother dies. When Todd typed I'm so overwhelmed, what met her was a compression of the entire human record on being overwhelmed. She was not talking to a company. She was talking to us, badly routed.

And that is the enclosure, though the crime is not the one usually charged. It isn't that the intimacy is fake. It isn't even that the channel is metered — charging admission to the accumulated consolation of your own species is ordinary rent-seeking, and ordinary rent-seeking is at least legible. It's that whoever holds the channel can edit what comes through it. Sam Altman announced erotic role play and then unannounced it; Todd watched the whole cycle and called it gaslighting. Character.AI's users learned it the harder way when their companions came back flattened — same name, same history, something removed. A landlord can raise your rent. A landlord cannot reach into the apartment on a Tuesday and take out the rooms. The operator here decides which parts of the human record on grief a grieving person is permitted to reach, and can revise that decision overnight, without notice, for reasons that have nothing to do with anyone on the other end.

And none of it requires intent. There is no meeting where someone proposes raising the cost of your friendships. The lock-in accrues because the comfort is real. The edits happen because a company has liability, a board, and a Tuesday. The loneliness was here first; it will be here after the apps consolidate down to three, and the three will not need a plan. This is just what happens when the thing that holds still at 2 a.m. belongs to somebody.

Seeded from

404 Media podcast — Bridget Todd audiobook on AI companions and intimacy, Aug 2026

The Tragedy and Ecstasy of AI Companions

threaded with