If your messages sound like a chatbot wrote them, it is not because the words are wrong. It is because AI-generated text carries a specific, identifiable set of habits: over-explaining, symmetrical sentences, relentless agreeableness, a question at the end of every message, punctuation so complete it reads like a work email. Every one is fixable in a single pass, and the fix is almost always deletion rather than addition.
Worth saying up front: the problem is not using AI. It is sending the first draft.
The tells
Over-explaining and stacked clauses
The model wants to be complete. Texting rewards being incomplete.
Bad: "That's really interesting that you're into climbing, I've always wanted to try it but never really had the chance since I moved cities and didn't know where to start."
Fixed: "i've wanted to try climbing for years. never knew where to start."
The symmetrical "not X, but Y"
Models love balanced constructions. Real people rarely build a text out of two matched halves.
Bad: "It's not that I don't like hiking, it's just that I prefer swimming."
Fixed: "i'd swim over hike every time."
The second has a preference in it. The first has a structure.
The rhythm of the em dash and the rule of three
Three items in a list. A dash for effect. Then a summary clause. It reads polished, which is the problem, because nobody texting at 11pm is polished.
Bad: "Your weekend sounds great, relaxing, productive, and social all at once."
Fixed: "how do you get a productive weekend AND a social one. teach me"
Relentless positivity
Every AI draft agrees with her. Everything is "so cool" or "amazing". Agreement is cheap and she has read it a thousand times.
Bad: "That sounds like such a fun adventure, I love that you're so spontaneous!"
Fixed: "booking a flight the night before is unhinged behaviour and i respect it"
Same warmth, but there is a person behind it now.
Interview mode
The model ends every message with a question, because a question is safe. Do it four times running and you have conducted an interview. This is one of the most common reasons a chat flattens out, seen from the other end in what to text when she stops replying.
Bad: "That's awesome that you went to Lisbon! What was your favourite part?"
Fixed: "lisbon in august is a bold choice. i lasted three days in that heat"
Perfect punctuation and capitalisation
Real texting is lowercase, clipped, comfortable with a fragment. A full stop on every line reads cold. A semicolon is a confession.
Bad: "I would love to hear more about that. Let me know when you are free."
Fixed: "tell me more. when are you free"
Generic specificity
The strangest tell. AI produces sentences shaped like details that contain none. "That sounds like such a fun adventure." Which part.
Bad: "Your job sounds so rewarding and interesting!"
Fixed: "wait, you do the night shifts? what time do you actually sleep"
Specificity requires having read her message. That is the whole trick.
Hedging
"I'd love to hear more about..." and "I was just wondering if maybe..." pre-apologise for the message. Cut the wind-up.
Bad: "I'd love to hear more about your trip if you feel like sharing!"
Fixed: "ok tell me about the trip"
Reformulating her message back at her
The model restates what she said before responding, because that is how assistants signal comprehension. In a text it reads like a support transcript.
Bad: "So you moved to Berlin for work and ended up staying. That's a big change! How are you finding it?"
Fixed: "you went for the job and stayed. what got you"
Why this matters more than style
Here is the part most advice skips. An AI-flat message is not just bland, it is structurally empty. No risk, no opinion, no self. Nothing she can disagree with or be surprised by. It gives her nothing to react to, so she reacts with the same nothing back, and two people trade pleasantries until one quietly stops.
Attraction needs a point of view. Not aggression, not negging, not manufactured tension. Just evidence that someone specific is on the other end of the phone. "Booking a flight the night before is unhinged behaviour and i respect it" hands her something to answer. "That sounds like such a fun adventure" hands her a receipt.
This is also why "just be positive" is bad advice. Relationship research has spent decades on responsiveness, the sense of being heard and understood, as a driver of closeness. Responsiveness is not approval. Being heard means someone engaged with the specific thing you said. Blanket agreement proves nobody read closely enough to react.
The test: if she could reply "haha yeah" and nothing would be lost, rewrite it.
The 20-second rewrite
Run this pass on any AI draft before it leaves your phone.
- Cut the first clause. The opening is almost always a run-up. Delete through the first comma and check whether it still works. It usually improves.
- Delete one adjective. The most enthusiastic one.
- Take a side. Find the sentence with no opinion in it and put one there.
- Drop the closing question if you asked one last message. Two running is an interview. Make a statement and let her carry it.
- Match her message length. She sent seven words, send seven. Four lines to her one signals imbalanced investment.
- Lowercase it. Kill the trailing full stop.
Six edits, all subtractive. Being interesting is mostly removing the parts that were not.
Make it sound like you, not like a press release
If you are using AI to draft, calibration is the whole job. A good tool should sound like your best texting day, not a brand voice. Feed it your register: how long your messages run, whether you punctuate, how you land a joke.
Then match hers. Someone who sends three-word replies and someone who sends paragraphs want different things, and one draft cannot serve both. This matters most when signals are scarce, which is why long distance texting and Instagram DMs to a stranger are the hardest cases: tone carries the entire relationship.
Wingyy's Translate feature exists for this. You type the raw thought, the ugly version, the thing you actually mean, and it returns a message in your voice rather than a model's. It reads the chat on screen first, so it has her register too. You still read, edit and send it yourself. How that compares to screenshot-and-paste is covered in the AI wingman app comparison.
Have a look at what Wingyy does, or start on the free trial and run the rewrite on whatever it hands you.
Common questions
Can people actually tell if a text was written by AI?
Usually not word by word, but they notice the pattern. Nobody thinks "this is GPT output". They think "this is a bit boring" and reply with less energy. The detection is emotional rather than forensic.
Should I use AI to text at all?
It is a reasonable tool if you treat it as a first draft rather than a final answer. Use it to break a blank page, to catch yourself interviewing someone, or to say a thing you already mean more clearly. Do not use it to invent a personality you cannot sustain in person. That bill comes due on the first date.
How do I make ChatGPT sound more like me?
Give it samples. Paste five of your own real messages and tell it to match the length, punctuation and rhythm, not just the topic. Then add constraints: lowercase, no closing question, under fifteen words, no exclamation marks. "Make it casual" is a generic prompt and gives you a generic voice.
What is the fastest fix for a robotic text?
Delete the first clause and the closing question. Those two edits remove the wind-up and the interview, the loudest tells. If you have another five seconds, lowercase it and cut the most enthusiastic adjective.