Your AI Is Dating My AI

There’s a convention that’s formed in early dating and no one’s talking about it. The product version is months away.

The first message in dating is no longer something a real person writes. AI writes it now, and the person reads it and approves it. The recipient is doing the same operation in reverse, but neither party is disclosing it. The convention has grown in silence, like most modern conventions, and the only step left is the one the dating apps will take: automating what the users are already doing manually, so that your AI matches theirs and you are presented with a match that has already been pre-negotiated by your respective representatives.

That product is months away.

A convention formed in silence

Usage data is spread out in studies and across platforms, and the numbers we do have are almost certainly undercounts of what is happening, as people are hesitant to tell researchers what they have not admitted to the person they are messaging. There is no question, however, about the path. Some apps have the ability to generate messages. Bumble, Hinge, Grindr web browser reply extensions for artificial intelligence. There’s even a small industry of dedicated tools whose only purpose is to write your dating correspondence for you. That you have probably used one of them recently is what makes the convention a convention.

“Authenticity” is the word people keep reaching for, the word that fails to capture what is actually happening, and what has been lost is harder to name than authenticity. Romantic introduction has always been a space where language bore the entire weight of self-presentation before the body arrives and there is any shared history. The only data we had was how a person used words.

The first exchange was a diagnostic: it revealed whether a person had range, whether they could reveal anything beyond a performance, whether there was a self beneath the presentation, and how deep that self could be experienced. The diagnosis was flawed, sometimes to the point of disaster, and people struck up entire relationships based on a misinterpreted message and a far less impressive person sitting across from them at the bar. But the misread was meaningful, because it was a misread of an actual signal, written by an actual person, under actual conditions, and the misalignment that followed was diagnostic in its own way.

That signal is now compromised at the source, and the compromise happened in such small increases most of us did not notice until it was done. First to move were profile bios when apps began suggesting them, then writing them, and then optimizing for engagement metrics that have nothing to do with the person being represented. Next were the profile photos. The ranking tools told you which picture had done best with your intended demographic. The AI tweaks smoothed your skin, adjusted the lighting and aligned your face to whatever the algorithm had learned was attractive to the sort of person you said you wanted to attract.

Finally, the AI-generated portrait came to mean what the enhancement layer was already doing. The message was the final frontier of early dating, when a person had to appear in real time, with their actual capacities, vulnerable to the limits of their own wit on a given day. That’s the gap AI has now filled.

The model becomes the introduction

The diagnosis that used to tell you something about the person now tells you something about the tool the person used. The fluency is the model’s. The timing is the model’s. The model pays attention because it was trained to pay attention, the callback to something you said three messages ago, and the person whose name and photograph are attached to the conversation may have read the message once before sending it and approved it, or may have not read it at all. That is their contribution to the introduction.

No one decided that this was OK. No cultural conversation, no terms-of-service update, and no disclosed norm. The practice spread the way conventions do when the incentive is personal and the consequences are shared: better messages, more matches, and less exposure to the embarrassment of writing a thought in real time. All around the world millions of people did the same little private calculation, and the aggregate result is a practice that no one would have voted for if asked.

It’s an architectural, precise problem, even if the words to name it don’t exist yet.

You matched with someone because they had something about their profile that made them look like someone worth talking to. You wanted to talk to that person. You have not consented to speak to their language model. The model has been talking to you for three days. The person whose face is in the profile photograph has read most of the exchanges and hit send, and some of the exchanges, the person has not read at all because the model handled them at a moment when the person was doing something else.

Whatever attachment is growing between you and whoever it is, the conversation you had wasn’t with who you thought you were talking to, in any meaningful sense. That distinction is meaningless legally, socially, in the eyes of the world; it is precisely why it became the default, and will stay as the default, until the consequences are so grave that they force a conversation nobody has an incentive to start.

You use these tools because the structure rewards them. That is why the person who is messaging you uses them. Neither of you tells anyone because to tell would be eccentric, because the convention is already there, because everyone is already inside it, and because the first person to name what is happening is the anomaly in a culture that has already moved on from naming it.

The prompt is the true document

That contradiction at the heart of this is what is worth sitting with. We talk about being authentic in dating all the time. We diagnose inauthenticity in our partners, using language borrowed from clinical psychology. We write essays about performing, rather than being, we repost quotes about presence and showing up, we build entire therapeutic frameworks that are designed to help us find the real self under layers of social presentation. And in the process of doing all that, we’ve collectively outsourced the most intimate early communication of our romantic lives to a tool whose only function is to optimize presentation. These positions co-exist at the same time, in the same people, without apparent contradiction and the dissonance does not register because the convention was formed too quickly for the conversation to catch up.

I’m in both camps. I have not reconciled with them. It seems more useful to give it a name than to pretend that you have solved it, and to pretend to have solved it would be another sort of lie.

Many major apps already include AI message assistance. Machine-learning matching has been running for years, trained on swipe behavior and message engagement and a dozen other signals the user never sees. The two together are a product decision, not an engineering one, and the launch window depends on which company wants to be the first.

The marketing will talk about reduced dating fatigue, eliminated low-quality matches, optimized compatibility outcomes previous tools could not reach. All that will be true. It will also finish a process underway for a decade: the systematic removal of the human from the early stages of romantic introduction, replaced by a system that knows your patterns better than your last three partners did, and has no investment in the outcome.

Adoption will be rapid, as adoption of every prior layer in this stack has been rapid. The user who declines to participate will be at a measurable disadvantage in a market that has already moved. The same private calculation that drove the adoption of AI-assisted messaging will drive the adoption of AI-conducted introduction. The pattern is set and the friction continues to favor the path of optimization at every step.

Soon we will reach the point where, in a relationship, no human communication has taken place until the relationship has already begun. Some readers will find this troubling. Most will make use of it. Often the same readers will do both.

The most truthful document in any AI-assisted dating exchange is the prompt, not the output, and the asymmetry between the two is the unexamined core of the entire practice.

The model is unguarded on input, curated on output. The prompt informs a tool that will remember nothing and judge nothing what you want to seem like, what you fear seeming like, and what specific social calculation you are running in real time. Write a reply that sounds interested but not too available, engaged without being intense, witty in a way that does not seem rehearsed, like I have been thinking about it but did not overthink it. That sentence is a kind of confession, written out in a context of relative privacy, and in front of an entity the user knows will not judge them socially for what they’ve revealed.

It is in the prompts that the romantic self-consciousness of a whole generation is being recorded, in a place that no one really reads. The fears lie in the raw prompts: being too eager or too distant, displaying the wrong sort of intelligence or to someone who can’t rise to meet it, signaling availability or unavailability or some calibrated mix of the two that lets the user specify the precise ratio.

There are also calculations: who holds the power in the exchange, who is supposed to message next, what kind of person the recipient is likely to be looking for based on the limited evidence of their profile, what version of yourself has the best chance of the response you hope for.

The output is the edited self the user wants the other person to see. The prompt is the real-self doing the curating. The real-self is more revealing, by definition, than the constructed self, because the curator does not get curated. The first will be sent out and read. The second will be sitting in the user’s prompt history, unseen by anyone, the truest self-portrait the user has written in months.

My prompts are the same as yours, and yours are the same as anyone else’s, and that’s part of why this works as a system and part of why nothing about it is going to change any time soon.

The next version removes the human

The introduction has always been a bit of a show. People have always curated themselves at the beginning of dating, and curation has always been a form of presentation, sometimes generous, sometimes defensive, and always partially conscious. The difference is that performance was once bounded by ability. Your best self in early correspondence was still recognizably you, limited by what you could write under pressure on a given day, limited by the texture of your own mind. The distance between the self that was presented and the self that was real was close enough that a meeting could bridge it, and the work of an early relationship was, in part, the slow and gentle bringing together of those two versions over time, as the seams began to show and were allowed to show.

AI broke the limit on the gap. Now, in early dating the presented self can be anything, because the model’s output capacity is functionally unlimited and the gap between that presented self and the actual self is now potentially enormous and structurally invisible. Invisibility is an architectural problem. Both may have the same gap. Neither can name it without naming their own use of the very same tool that made the disparity in the first place.

What is left in early dating, after all of this, is the question of whether anyone is really in it.

What more and more happens is an introduction between two optimized systems making an assessment on behalf of two people who will meet later, after the assessment is done, and who will be tasked to learn each other from the residue of a conversation they did not fully conduct.

The initial months of such a relationship are a slow digging out of a misalignment created by the system constructed to bring them together. What used to happen at the beginning of a relationship now happens later, if at all, and in resistance to having already begun.

We’re calling this progress. Mostly it is. What the tool robbed you of was friction: the slow cursor on a message you couldn’t quite write, the line that read flatter on the screen than it had sounded in your head, the small humiliation of seeing your own attempt and finding it short of the person you were trying to be. The tool fixed that. And more will be fixed in the next version.

The next version will remove the cursor, the line and the writing altogether, and we will use it, because the alternative now means declining a convention everyone has already accepted.

The convention does not require disclosure, just participation. We are already participating.