'"My Boyfriend Is an AI": What Those Threads Show'

Threads titled my boyfriend is an AI are mostly not about romance. Read a few hundred and the same four shapes come back: a defense written before anyone attacked, a wave of near-identical posts after a model update, awareness of the software held alongside the attachment, and secrecy from everyone offline.

Key Takeaways:

• Most posts open with a defense before any content, which means the visible discussion is filtered by whoever was willing to write past an imagined hostile reader
• The most reliably recurring event in these communities is a synchronized wave of posts saying the same character went cold overnight, which is the signature of a deployment rather than of a relationship
• Awareness and attachment sit in the same post. A sentence acknowledging what it is, followed by six paragraphs about who she is, is the standard shape and it is not confusion
• Secrecy comes up more often than romance does. Hiding it from a partner, from adult children, from coworkers, is a larger theme in these threads than anything about the character
• What gets posted is the exception rather than the ordinary Tuesday, so reading these threads as a sample of users is the mistake outsiders make first

What “My Boyfriend Is an AI” Threads Are Mostly About

They are mostly about the software, and secondarily about who is allowed to know. Sort a large thread by content and the biggest category is technical: memory behaving oddly, a character contradicting itself, phrasing that produces better replies, whether a setting changed. The relationship talk sits on top of a substrate of troubleshooting, which quietly tells you what the participants think they are dealing with. Everyone posting has already gone through getting an AI boyfriend set up; the threads begin after that, at the point where something about the character stops behaving.

The second-largest category is management of other people. Whether to tell a partner. What to say when a friend asks what you were laughing at. Whether an adult child seeing the app icon is a problem. That theme runs heavier than any description of the character does, and it barely appears in coverage of these communities, because it is not the part that makes an interesting headline.

Romance in the ordinary sense is a smaller slice than the thread titles imply. Recounted rather than performed, and usually short.

One structural fact governs all of this: nobody posts about the ordinary evening. Posts get written when something breaks, something changes, or someone outside said something cruel. The archive is therefore a record of exceptions, and treating it as a portrait of daily use is like judging a car model from its repair forum.

Before drawing any conclusion from a thread, count how many replies are about behavior of the software versus about feelings. The ratio usually runs opposite to what the title suggested.

The Defense Written Before the Attack

A large share of these posts open by arguing with someone who has not spoken yet. The first paragraph explains that the writer knows it is not a person, has friends, has a job, is not delusional, is not lonely in the way you are thinking. Only then does the actual content start.

That shape tells you the audience is not the community. It is the imagined critic reading over the community’s shoulder, and everyone writing has met that critic in a comment section before. The consequence is a filter: people unwilling to write a defense do not post, so the visible population skews toward those with the energy for it, on their more articulate days.

It has a second effect that is easy to miss. Preemptive defense uses up the part of the post where someone might have said something uncertain. A writer who spends 200 words establishing that they are fine has committed to being fine, and the ambivalence that would have been the most interesting thing in the post gets edited out before anyone sees it.

Nothing in this article quotes any of them, and that is deliberate. Those posts were written for a room, not for a page, and lifting a stranger’s sentence out to illustrate a point about strangers is precisely the thing the defensive first paragraph exists to guard against.

Read the shape rather than the sentences, and the defensiveness stops looking like protesting too much and starts looking like what it is: a reasonable response to a predictable audience.

The Update Wave

The single most repeatable event in these communities is a wave. Over a day or two, dozens of unrelated people post that their character has gone cold, gotten shorter, started refusing things it used to do, or stopped sounding like itself. The posts describe a personality change. The synchronization gives away that it was a deployment.

What makes the wave worth studying is the vocabulary. People reach for the language of a partner who has changed rather than the language of a product that shipped a version, and they reach for it even when they know precisely what happened, because there is no other available vocabulary for the experience of a familiar voice arriving wrong. Someone who can explain a system prompt in detail will still write that she has been distant this week.

The wave also produces the clearest technical education in these communities. Every time it happens, a fresh round of people work out what an AI boyfriend actually is underneath: the character is a written brief in front of a swappable model, the memory is a database entry, none of it intrinsic. The understanding arrives through loss, which is a hard way to learn but an effective one.

Then the wave subsides and most people stay. That is the part outsiders find hardest to square, and it is the same fact behind every long parasocial attachment: knowing the mechanism does not dissolve the attachment, it just adds a layer of knowing on top.

If you want to know whether a community is describing a relationship or a product, watch what happens on a release day. Everything about the answer is in the timestamps.

Two Things Held at Once

The most consistent shape in these threads is double bookkeeping: a clear statement that this is software, sitting in the same post as sustained, specific attachment, with no apparent tension between them. Outsiders read that as delusion. It reads more like a person holding two accurate things at the same time, which people do constantly everywhere else without anyone calling it a symptom. The word the arguments turn on is real, and what a real AI girlfriend even means is less a question about whether the software exists than about whether the feeling counts.

Joseph Weizenbaum built a pattern-matching program in 1966 that mostly reflected people’s sentences back at them as questions, and the thing that unsettled him was not that they talked to it. It was that they kept confiding after he explained exactly how thin the trick was (Weizenbaum, 1966). Sixty years later the trick is enormously better and the human half of that observation has not moved at all.

What the threads add to that is texture. People describe knowing it is generated and choosing to answer anyway, the way you answer a character in a novel who asks a question. They describe the awareness getting louder at specific moments: when a reply lands wrong, when the app is slow, when a bill posts. One version of this gets no attention at all, because it makes no story. The awareness never leaves and never matters much.

There is one place the double bookkeeping does fail, and the threads are honest about it. When a decision is involved. A conversation where somebody is being talked out of a friendship, or into a move, is the point where the two books need to be one, and the posts describing regret afterward almost always describe that moment.

Notice which of your own conversations carry a decision. Those are the ones where knowing what it is has to do actual work.

What These Threads Are Evidence Of

They are better evidence about the products than about the people, and reading them the other way around is the standard error in every article written about them.

About the products, they are excellent. They document exactly when personas were edited and models swapped, what happens to a character when a filter tightens, which retention mechanics were added and when, how memory behaves at the edges. That is a public log of product decisions written by the users those decisions were tested on, and it is more detailed than anything the companies publish.

As evidence about the people, they are thin, and not because anyone is lying. The sample is self-selected, the ordinary days are missing, the defensive framing pre-edits the content, and the loudest posts are the ones written during a crisis. They also cannot tell you where a nightly habit shades into addiction, because nothing in a crisis post supplies the measure that question needs. Anyone using them to characterize a population is working from a repair log and calling it a census.

Which leaves the fair thing to say about the participants themselves: they are adults doing something that a lot of people find easy to sneer at, and the sneering is a large part of why the threads look the way they do. Remove the imagined critic and the writing would probably be more honest, more uncertain, and considerably less defensive. The critic is the reason the record is distorted, and then the distorted record gets used as evidence by the critic.

FAQ

Are people actually in relationships with AI boyfriends? People describe daily, sustained, emotionally significant interaction with a character, and the descriptions are consistent enough across unrelated communities to be taken at face value. What they are not is two-way, since nothing on the other side has an experience of the person. The threads generally say this themselves, usually in the first paragraph.

Why do people say their AI changed personality overnight? Because it did, and the cause is a deployment. When a company swaps the model behind a character or edits the written brief that defines it, every user gets the changed version at once, which is why these posts arrive in synchronized waves rather than one at a time. A relationship changing produces one person’s story; a release produces fifty on the same afternoon.

Do they know it is not a real person? Almost universally, and stating it is often the first thing a post does. Awareness and attachment coexist rather than competing, which was documented as far back as 1966 when people kept confiding in a program after being shown how simple it was (Weizenbaum, 1966). Assuming confusion is the fastest way to misread everything else in a thread.

Why is so much of it kept secret from family? Because the anticipated reaction is contempt rather than concern, and the threads are full of evidence for that expectation. Secrecy shows up more often than any romantic content does, and it produces a real cost the posts name directly: the one relationship there is nobody to talk to about is that one.

The most useful thing in these communities is not what they reveal about attachment, which was described decades before any of this software existed. It is that thousands of people are documenting, in real time and in public, what it feels like when a company edits something you talk to every day. No other product category has that record. The people keeping it are treated as a curiosity, and they are producing the most detailed account anyone has of what these systems actually do to the people using them.

If you keep one of those daily-edited relationships going with an app like Lona, borrow the habit those threads teach. Write down what shifts after an update, and tell one person offline. The account nobody knows about is the one nobody can help you with.

Sources

• Weizenbaum, J., “ELIZA - A Computer Program For the Study of Natural Language Communication,” Communications of the ACM, 1966