Does AI make your texts sound like a robot?

Usually, yes. And for the first time there is proper research on why. When an AI model rewrites a message, the meaning mostly survives. The way you say things does not. Contractions disappear, “I” turns up less often, the small words that carry your rhythm get tidied away, and what comes back reads like it was written for anyone.

I co-founded Subtext, an app built around exactly this problem, so weigh what I say about it accordingly. What I can offer is that I went through the studies properly, and I have linked every one so you can check me. This is what they show, what they do not, and what you can do about it, whether you use our app or paste your drafts into ChatGPT.

What happens to a message when AI rewrites it

The largest study so far came out in Nature Human Behaviour in August 2026. The researchers looked at more than 880,000 texts, including controlled experiments where the same pieces of writing were run through several AI models with prompts like “polish”, “make it clearer” and “make it more natural”. The core content stayed put. The variety in how people wrote did not. Across their datasets and models, the spread of writing complexity shrank by 21 to 50 percent, and classifiers that guess a writer’s gender, age, politics or values from their writing got noticeably worse on the rewrites, with the errors leaning toward the majority profile (The shrinking landscape of linguistic diversity in the age of large language models)¹.

That study’s rewriting experiment used Reddit posts and academic abstracts. A smaller 2026 study, still a preprint, looked at something closer to a text message: 300 personal stories, each rewritten by three leading models. The pattern was the same every time. Fewer contractions. Fewer first-person pronouns. Fewer of the little function words. Longer words and fancier punctuation. It did not matter whether the instruction was “improve this” or just “rewrite this”. Telling the model to preserve the writer’s voice cut the effect by about a third, but the direction never changed, and the model was worse at keeping what it likes to remove than at holding back what it likes to add (Voice Under Revision: Large Language Models and the Normalization of Personal Narrative)².

Two findings pull the other way, and they matter. A 2025 study of news articles from before and after ChatGPT found more AI-typical words but no measurable loss of vocabulary variety on the metrics it used, so the effect is clearest in controlled rewriting and less clear in the wild (Testing English News Articles for Lexical Homogenization Due to Widespread Use of Large Language Models)³. And a peer-reviewed experiment on co-writing found that when essays became more similar, the sameness came from the text the model inserted. The writers’ own contributions stayed as varied as before (Does Writing with Language Models Reduce Content Diversity?)⁴.

That second finding is the useful one. The damage comes from replacing your sentences, not from small edits to them.

Why your meaning survives and your voice does not

Researchers who work on writing style treat two things as separate. One is what a message says. The other is how it is said, and “how” is made of very ordinary material: which everyday words you reach for, how long your sentences run and how much they vary, your punctuation habits, the hedges like “kind of” and “honestly”, whether you write “I’m” or “I am”, the slang, the emoji, the switch into another language mid-sentence. A 2021 study of online writing found these habits are consistent enough that a model trained on tens of thousands of writers can usually tell, from two short posts, whether the same person wrote both (Idiosyncratic but not Arbitrary: Learning Idiolects in Online Registers Reveals Distinctive yet Consistent Individual Styles)⁵.

The meaning needs none of those features. So a rewrite can keep the meaning perfectly and lose all of them. That is exactly what the Nature study measured. What was said stayed highly similar, while the variety in how it was said collapsed.

That makes “same message” and “same voice” two different promises. Most tools only make the first one, and most people only notice the second is missing after they have hit send. I wrote up what the big three chatbots each do to your words separately. Here, the question that decides whether any of this matters is the next one.

Does the other person actually notice?

The evidence here is more specific than I expected.

A 2026 study published at CHI, the main conference for research on how people use technology, tested exactly this in romantic relationships. In a randomised experiment with 704 people reading messages from a partner, the researchers varied two things: whether the AI had done a light tone rewrite or drafted the whole message, and whether the sender admitted the AI helped. Heavier AI drafting reliably made the message feel less like the sender’s own and less authentic, and being clearer or more polished did not make up for it. The striking result was about apologies. When the sender said “I had AI help with the tone” on a lightly edited apology, readers found it more authentic and were more forgiving. When the sender said the same thing about a fully AI-drafted apology, both slipped. Messages with more of the sender’s own quirks scored higher on authenticity throughout (Is It Still You? Attributing Authorship and Authenticity in AI-Assisted Romantic Communication)⁶.

A second study, from 2025, explains part of why. When 399 people read messages labelled as AI-assisted, the label did not make the sender look bad, but it did make the message stop counting as evidence. An unassisted apology read as warm. An unassisted blaming message read as cold. Add the AI label and both moved toward neutral, as if the message said nothing about the person at all (Explaining the Reputational Risks of AI-Mediated Communication)⁷.

Disclosure has a cost of its own. Across 13 preregistered experiments, people who said they had used AI were trusted less than people who did not, in settings from grading to design work (The Transparency Dilemma: How AI Disclosure Erodes Trust)⁸. Though in emotionally neutral, one-off situations like a trust game with a stranger, AI help made almost no difference to trust whether disclosed or not (Writing with AI boosts trust-building efficiency)⁹.

Read together, the pattern is consistent. The more personal the message, the more a full rewrite costs you, and the more a light touch on your own words can help. The apology research in particular deserves its own piece, and I gave it one: how do you apologise over text?

The flattening hits some people harder

If you write in a second language, or you move between languages when you text, you will have felt this already.

In a 2025 experiment, 118 people in India and the United States wrote short essays about their own lives, just over half of them with AI autocomplete switched on. With suggestions running, the Indian writers drifted toward the American group’s patterns, and started describing their own food and festivals from a Western point of view (AI Suggestions Homogenize Writing Toward Western Styles and Diminish Cultural Nuances)¹⁰.

That is one study, in English, and it cannot speak for every language. But the mechanism is easy to see. The things a model is most likely to “correct” in your text, the borrowed word, the phrasing that came from your first language, the switch to Turkish or Spanish for the part that only makes sense in Turkish or Spanish, are the things that are most you.

Can an app learn how you write?

Partly. And you should be suspicious of anyone who says more than that, including me.

The most thorough test so far gave leading models from OpenAI, Google, Meta and DeepSeek real writing samples from more than 400 people and asked them to write new text in each person’s style. Examples helped a lot. In structured writing like email, an authorship checker accepted the imitation as the real person’s writing over 95 percent of the time. In casual writing like blog posts, it was fooled only 17 to 21 percent of the time, adding more examples barely moved that, and the output drifted back to a generic tone (Catch Me If You Can? Not Yet: LLMs Still Struggle to Imitate the Implicit Writing Styles of Everyday Authors)¹¹. A separate benchmark from April 2026, still a preprint, scored four personalisation methods against how similar real people are to themselves. Every method scored below the level at which two different people resemble each other (Theory-Grounded Evaluation Exposes the Authorship Gap in LLM Personalization)¹².

So no tool can honestly promise you will sound exactly like yourself. What a tool can do is narrower and more useful. It can change as little as possible, work from your words instead of replacing them, show you what it changed, and let you decide.

That is how we built Subtext. Subtext reads the thread and your draft, and before it rewrites anything it tells you how the message is likely to land and which of your words are doing it. Then it offers three versions. The first stays close to what you wrote, the last is a fuller rephrasing, and each one says how much of your wording it changed, so you can see at a glance whether you are looking at a polish or a replacement. You can nudge any of them warmer or shorter, or open Manual edit and change a word yourself before you copy it. The default is the light touch, since that is what the research says survives contact with the person reading. Try Subtext in your browser

How to keep your voice, whatever tool you use

These work in any app, ChatGPT included.

1. Ask for the smallest job. “Fix grammar and typos only” gets you a very different result from “improve this”.
2. Say what to keep, out loud. Contractions, “I”, the emoji, the swearing, the sentence fragment. Models are worse at protecting what they would normally remove than at holding back what they would add, so name it.
3. Never accept a whole rewrite unread. Compare it with your draft line by line and take only the lines that are still you.
4. For anything emotional, keep your own draft as the base. In the romantic-communication study, an apology with a light AI tone edit landed better than one the AI drafted in full.
5. Read it out loud before sending. If you would not say it to their face in those words, do not send it.

If you want the same check done for you, the piece on what an AI message checker can check, and how to do it yourself walks through it, and I have compared the AI message rewriters on the market by whether they diagnose before they rewrite.

Questions people ask

What is the best app for rewording a message without making it sound like a robot wrote it? The one that changes the least and tells you how much it changed. On the evidence above, a tool that replaces your sentences will flatten your voice whatever its marketing says, so look for a version that stays close to your draft, a label for how far each version has moved, and a rewrite you can decline. We built Subtext that way, and the checklist above works in any app.

Is there an app that rewrites my texts in my own voice, not in that generic AI way? Partly. No current method matches how consistent a real person is with themselves, so treat “in your own voice” as a claim to test rather than a feature to trust. The practical version is a tool that starts from your draft, keeps your contractions, hedges and emoji, and labels how far each version has moved from your wording. That is what I would look for, and what Subtext does.

Is there an app that rewrites my message but still sounds like me? I’d put it as “closer than most, never exactly”. The research above says every current method falls short of how consistent a real person is with themselves, so any app that promises a rewrite that is indistinguishably you is overpromising. What you can get is a rewrite that starts from your sentences, keeps the habits that carry your voice, and tells you how far it moved. Subtext works that way, and its first version is deliberately the one closest to what you typed.

How do I make my texts sound warmer without losing my own voice? Add warmth in your own words rather than asking for a warm rewrite. A full rewrite is what strips your voice. Name one specific thing about the person or the moment, keep your usual phrasing around it, and if you use a tool, ask it to change only the line that reads cold, not the whole message.

Why does ChatGPT make my messages sound so formal? Because it is trained to produce fluent, average text, and average English is more formal than how anyone texts. Rewrites push toward longer words, expanded contractions and fewer first-person pronouns even when nobody asked for formality. If you want casual, you have to ask for casual, and you have to name the specific things you want left alone.

Should I tell someone I used AI to write a message? It depends on how much AI did. In the romantic-communication study, admitting to a light tone edit on an apology made it land better. Admitting to a full AI draft made it land worse. If the words are yours and the AI tidied them, saying so is fine and can even help. If the AI wrote it, the honest move is to rewrite it yourself first.

Is this the same for texting in German, Turkish or Spanish? Most of the research is in English, so I cannot say for certain. The one cross-cultural study found AI suggestions pulled writers toward American patterns. If you write in another language or switch between languages, the safest habit is to tell the tool explicitly which words and phrases are off limits.

Sources

Every link above goes to the primary source, numbered in order of appearance. Preprints are marked as such; treat their findings as provisional until peer review.

1. Sourati, Z., Karimi-Malekabadi, F., Ozcan, M., McDaniel, C., Ziabari, A., Trager, J., Tak, A. N., Chen, M., Morstatter, F. and Dehghani, M. (2026). The shrinking landscape of linguistic diversity in the age of large language models. Nature Human Behaviour.
2. van Nuenen, T. (2026). Voice Under Revision: Large Language Models and the Normalization of Personal Narrative. arXiv preprint 2604.22142.
3. Fitterer, S., Gangl, D. and Ulbrich, J. (2025). Testing English News Articles for Lexical Homogenization Due to Widespread Use of Large Language Models. ACL 2025 Student Research Workshop.
4. Padmakumar, V. and He, H. (2024). Does Writing with Language Models Reduce Content Diversity? ICLR 2024.
5. Zhu, J. and Jurgens, D. (2021). Idiosyncratic but not Arbitrary: Learning Idiolects in Online Registers Reveals Distinctive yet Consistent Individual Styles. EMNLP 2021.
6. Fan, G., Liu, D. and Pan, L. (2026). Is It Still You? Attributing Authorship and Authenticity in AI-Assisted Romantic Communication. CHI 2026.
7. Khadpe, P., Wenzel, K., Loewenstein, G. and Kaufman, G. (2025). Explaining the Reputational Risks of AI-Mediated Communication: Messages Labeled as AI-Assisted Are Viewed as Less Diagnostic of the Sender’s Moral Character. AIES 2025.
8. Schilke, O. and Reimann, M. (2025). The Transparency Dilemma: How AI Disclosure Erodes Trust. Organizational Behavior and Human Decision Processes, 188.
9. Purcell, Z. A., Jakesch, M., Dong, M., Nussberger, A.-M. and Köbis, N. (2025). Writing with AI boosts trust-building efficiency. iScience, 28.
10. Agarwal, D., Naaman, M. and Vashistha, A. (2025). AI Suggestions Homogenize Writing Toward Western Styles and Diminish Cultural Nuances. CHI 2025.
11. Wang, Z., Tripto, N. I., Park, S., Li, Z. and Zhou, J. (2025). Catch Me If You Can? Not Yet: LLMs Still Struggle to Imitate the Implicit Writing Styles of Everyday Authors. Findings of EMNLP 2025.
12. Sawant, Y. G. (2026). Theory-Grounded Evaluation Exposes the Authorship Gap in LLM Personalization. arXiv preprint 2604.26460, accepted at the ICML 2026 CTB workshop.