Should I use AI to write my dating app messages?

For a nudge, yes; for the whole conversation, no. AI is good at getting you unstuck — reading a profile you can't find a hook in, or unsticking a thread that died. It is bad at being you for three messages and then handing the conversation back, because the change in voice is visible from the inside of the conversation and it arrives exactly when she is deciding whether to meet you.

Do AI-generated openers actually get more replies?

They did, and that is fading. An opener works because it is specific to her profile and to you; a tool trained on what has worked produces what everybody else's tool produces. She is not comparing your message to silence — she is comparing it to the other AI-written openers in her inbox that week, and sameness reads as a mass send.

What is the risk of an AI wingman app?

The handoff. Messages one to three are written by a tool and message four is written by you, and the gap between them is the thing she notices. The same problem as a photo that flatters you: it works right up until the moment it has to survive contact with the actual person.

The AI Wrote a Great First Message. Then She Met You.

A phone on a table showing an open dating-app conversation

The pitch on these apps is direct, and it is the same pitch everywhere:

"Get more matches, better conversations, and actual dates."

Upload screenshots of your Tinder or Hinge profile, let the AI analyse it, and let it write your replies for you.

The first part of that is a real product. The second part has a problem nobody markets, and it is not the one you would expect. The messages are usually fine. The trouble starts when you take over.

The handoff

Picture how it actually goes.

Message one is written by the tool. It is specific, it references something in her profile, it has a light question at the end. She replies. Message two and three are also the tool, and they are also good — this is not hard for a language model, it is close to the easiest thing a language model does.

Somewhere around message four you stop pasting. Maybe the app ran out of free generations, maybe you are on the train, maybe it simply felt silly. And your actual voice arrives in the middle of a conversation that has been performing a different one for three messages.

She does not consciously think "this is a different person". She thinks the conversation went flat. Something loosened. The version of you that opened is not the version of you that is here now, and the drop-off registers somewhere below the level of an articulated thought — which is exactly where attraction is decided.

The handoff is the product's real output. Not the opener.

It is the same failure as a photo that is not you

This site has written a lot about photographs that improve you. A body you do not have, an age you are not, a face the model quietly averaged. The argument every time is the same: the photo has one job, and that job is checked in person.

Outsourced text is that argument, moved to the conversation.

A first message that was not written by you gets you to message four. A photograph that was not you gets you to the coffee shop. In both cases the thing you bought did its job right up to the point where it had to survive contact with the actual person — and in both cases the bill is paid at the most expensive moment possible, when she has already invested and is now revising downward.

The earlier a mismatch shows up, the cheaper it is. A flat opener costs you a reply. A great opener attached to a different person costs you a date and some of her good faith.

The part that is quietly getting worse

Set the handoff aside for a moment, because there is a second problem and it is on a timer.

These tools work by producing what has worked. That is the whole mechanism — trained on successful openers, tuned toward the ones that got replies. Which means every user of every one of these apps is being steered toward the same small set of messages.

She is not comparing your opener to nothing. She is comparing it to the eleven other openers in her inbox this week, and if three of them opened with the same warm, specific, lightly-teasing question about her hiking photo, all three of you read as automated. Not one of you individually — all three, retroactively, including the one who wrote it himself.

This is the identical mechanism that makes AI photos detectable. A model's output converges toward an average, and once enough people are standing on that average it stops being attractive and starts being a signal. Openers are a few years behind photographs on the same curve.

What AI is genuinely good at here

Not nothing. The useful version is narrow and it is worth naming, because "don't use AI" is not honest advice in 2026.

Getting unstuck. You have read her profile four times and there is no hook in it. Asking a model for five angles, then writing your own message from the one that sparked something, is a completely different act from pasting its output. You used it to think, not to speak.

Diagnosing a dead thread. Why did this conversation die? is a question a model can answer usefully, after the fact, without ever writing in your voice.

Fixing your bio. This is the strongest case and almost nobody uses it. A bio is written once, revised slowly, and is genuinely supposed to be a polished version of you. There is no handoff problem, because there is no moment where you take over.

The pattern: use it on the parts that are written once, not on the parts that are a live conversation. A bio can be edited. A conversation cannot be edited, only continued.

The test, moved to text

We keep coming back to one standard for photographs: would somebody who knows you recognise you walking into a coffee shop.

The message version is simpler than it sounds. Read the opener out loud. If you would not say it in that order, with those words, to a person standing in front of you, it is not going to hold the conversation it opens.

That is not a rule against being polished. Write it three times, cut the boring half, make it better than your first instinct — that is all fine, and it is what a good bio is. It is a rule against being someone else, for three messages, at the exact moment when the only thing being assessed is who you are.

What we do

We make photographs, and that is the whole product. We do not write your messages, we do not analyse your chat history and we do not sell an opener generator. That is a deliberate boundary rather than a missing feature.

The reason is the argument above. The photograph is the part that can be honestly improved without becoming someone else — better light, a better scene, a version of your own face that is not a bathroom selfie at 11pm. The conversation is not that. There is no lighting equivalent for a first message; there is only you writing it or somebody else writing it.

So the instruction that goes with every generation is about holding a person still — do not beautify, redesign, age-shift, slim, widen, or change the face — and the conversation afterwards is left to you, because it was never ours to hold.

The honest limit: better photographs get you more matches, and more matches means more conversations you then have to actually have. We do not solve that half, and a tool that claims to is claiming to be you.

The short version

AI writes a good first message and a good second one. The failure is the handoff — message four is you, and the change in voice lands exactly when she is deciding whether to meet.

Use it to get unstuck, to diagnose a dead thread, or to fix a bio you write once. Do not use it to hold a live conversation, and read anything you are about to send out loud first: a message you would not say is a message you will have to explain later.