What's the best AI writing assistant for text messages?

For messages between people, Subtext. For essays, reports and marketing copy, something else on this page. The rest of this piece explains why that distinction holds.

The general answer to “best AI writing assistant” is usually Grammarly or ChatGPT, and for documents that answer is right. Personal messages are a different problem. The thing that makes a text to your sister work is not correctness, and the research on AI help in personal messages points at a specific cost.

Before you read on, you should know that I co-founded Subtext, so this piece recommends my own product in its first sentence. I am not a neutral reviewer of this category and will not pretend to be one. What I can offer instead is an argument built on the published research about what correctness tools do to personal messages, with every study linked so you can check whether it says what I claim it says.

The problem every other AI writing tool created

There is a body of experimental research that should worry anyone building an AI writing tool, and it shaped almost every decision in how we built Subtext.

In 2023, Hohenstein and colleagues published a controlled study in Scientific Reports¹ on what happens when people use AI assistance in conversation. The efficiency findings were what I’d expect. Smart replies produced 10.2% more messages per minute. Conversations carried more positive emotional language. What I found uncomfortable was the second one. When a participant suspected their partner of using AI, they rated that partner as less cooperative and affiliated with them less, whether or not AI had been used.

Follow-up work sharpened the mechanism. Khadpe, Wenzel, Loewenstein and Kaufman, presenting at AIES 2025², describe it as signal diagnosticity. A message read as AI-assisted stops functioning as evidence about the sender’s character. Their finding is specific, and I’d call it damning for the category. An AI-assisted apology makes the sender appear less warm than if they had written it badly themselves. Earlier CHI work found the same pattern with Airbnb profiles³ and with condolence and support emails⁴, where trust dropped as the perceived author shifted from human toward machine.

So the market filled up with tools that generate polished text from nothing, and the polish can become the tell. In these studies the damage came from the reader believing AI was involved, whether or not it was. The cost lands on the sender, on precisely the dimension the tool promised to improve.

Subtext is built around this finding. Its homepage sums up the thesis in one sentence. You can tell when a message came from a robot, and so can they.

Starting from your draft

The architectural decision that follows from that thesis is the one that matters most, and it separates Subtext from every general-purpose model.

ChatGPT and Claude write a message for anyone, as I found comparing what the big three chatbots do to your words. You describe a situation, they synthesise prose. The output is competent and generic by construction, because there is no version of you in the input. Grammarly and Wordtune take the opposite approach and correct toward a standard, which strips the informal warmth and the slightly awkward specificity that make a personal message read as human.

Subtext takes what you already wrote and works on that. A half-finished draft is enough. The composer opens with “Write or ask me to,” and a fragment of an angry text is a valid starting point. The model has your words, your rhythm and your particular way of getting to the point, and it refines from there. The app then shows how much it changed, so you can pull any of it back.

The newest evidence points the same way, with a limit. In a 2026 CHI study of AI help with romantic messages⁵, with 152 and then 704 participants, the more of a message the AI drafted, the less it felt like the sender’s own. A light tone edit the sender owned up to made apologies land better, though the results differed for boundary requests, so even a light edit needs a read before you send it.

The review that keeps recurring across both stores says a version of the same thing. One App Store user put it as being able to put thoughts into words in a way that still feels like themselves, warmer and clearer. Another described it as texting with a wise friend. A Google Play reviewer with a ChatGPT Pro subscription called it wholly different, which is the comparison that counts to me.

Reading the message before writing it

Before Subtext suggests anything, it reads what you wrote and watches for the emotional triggers hiding in it.

Instead of a tone label, it marks the exact words that could land as harsh, passive, or ambiguous, and names the feeling underneath them. Here is a real example from the product, tags and all.

Real draft, tagged by Subtext “look i really need u to stop sending voice msgs without any text… it freaks me out and u know that but u keep doing it anyway?!!”

Passive-aggressive · carrying loaded language · driven by frustration

Each tag is tappable and explains itself.

Compare this to what the incumbents offer. Grammarly’s tone detector⁶, released in 2019 and available on mobile through its keyboard, returns adjectives like confident or polite. Grammarly has since added Reader Reactions, which simulates how a reader you pick, such as your manager or a client, might interpret something you wrote⁷ in Grammarly docs, and Copilot in Outlook coaches work email on tone, clarity and reader sentiment⁸. Both are aimed at work writing. Apple Intelligence Writing Tools⁹ offers rewrite, proofread, and a small set of tone presets. Google’s Magic Compose¹⁰ offers seven styles including Shakespeare and Lyrical, which tells you what it is for. The Apple and Google options are stylistic filters. Subtext is doing interpersonal risk analysis, and the difference shows up most when the message matters.

Bringing in the whole conversation

Subtext reads the whole thread. Most drafting tools see one message in isolation, which is the wrong unit, because a reply is a response to something.

You can paste a screenshot of the conversation from WhatsApp, iMessage, Slack, or a dating app, and it reads the whole exchange for context. You can record a voice note and talk through what you mean, no typing required. It handles their voice messages too, which solves the specific modern problem of a four-minute audio note you need to answer and cannot bring yourself to replay.

For messages carrying real weight, the app stops and asks whether you want to hold a boundary or reassure someone, then writes from your answer. In the voice-note example above, the two options offered are that the text still matters to you, or that you are fine and there are no hard feelings. Those produce different messages, and I don’t think any amount of prompt engineering would reliably get you to the right one.

Choosing between versions

Subtext’s interaction model deserves its own attention, because it removes prompt-writing entirely.

You get three directions side by side. Warm and clear, reassuring, short and kind, each tagged with how it reads. From the same voice-note situation, one version leads with “You didn’t do anything wrong,” another with “We’re completely fine,” a third with “All good.” Reading them against each other is a faster way to find out what you meant than describing it in advance to a chatbot.

From there, adjustment is a tap. Warmer, shorter, more casual, add humour, add confidence. Behind each one sits a considered instruction, more detailed than a keyword. The humour control specifies light humour to soften and avoids sarcasm entirely. The shorter control preserves the language you wrote in. These are the details that separate a product someone thought about from a wrapper around a model.

It works this way across 17+ languages, including German, Spanish, French, Turkish, Japanese, and Korean, and it writes in each language directly, without translating out of English. The tags, the explanations, and the safe-to-send note all switch with the language.

Privacy that matches the content

The category asks people to hand over their most sensitive correspondence to Subtext and its competitors alike: arguments with partners, resignations, messages to family members they are not speaking to. Most tools treat that data the way they treat any other input.

Subtext runs a zero-training policy. Drafts, transcribed audio, uploaded screenshots, and context answers are used to generate your options and nothing else. None of it enters a training corpus.

Storage is deliberately temporary. Unpinned conversations and their attachments delete themselves from the servers five days after you last touched them. Pin a thread and the window extends to 90 days, which is enough to keep it synced across your phone and the web client. Delete something yourself and it goes immediately, attachments included. Data moves under transport encryption, and the privacy policy¹¹ names every provider in the chain.

The app also does not send anything for you. You copy the version you want and send it from your own hands, which is a small design choice with a large implication about who is responsible for the message.

Where it runs

Subtext is built for the phone first, on iOS and Android, with a browser client at web.subtext.it that syncs to the same account and works across Safari, Chrome, Firefox, Edge, Opera, and Brave.

The web client matters more than it appears to. Apple Intelligence follows you around iOS and stops at the edge of it. Subtext follows the conversation, including the one happening in a browser tab on a work laptop.

Downloads have passed 100,000 across platforms, and the App Store¹² rating sits at 4.9 as of August 2026.

What it’s for

The clearest signal comes from what people report using Subtext for.

Neurodivergent users have become a meaningful part of the base, enough that there is a page on an AI tone checker for neurodivergent adults. One Google Play reviewer describes it as helping send messages without worrying about misunderstanding and tone, specifically because they are neurodivergent. Another describes long, ADHD-shaped texts being condensed while the substance survives. This is a use case the general-purpose models serve badly, because their instinct is to normalise the input rather than to clarify what was already there.

Beyond that, the pattern is consistent across contexts. Holding a boundary at work without sounding defensive. Cancelling plans honestly and kindly, which one reviewer named exactly. Answering a message from someone you are dating without reading it eleven times first. One user reported it saved them an argument, which I think is a fair description of the value.

The case, stated plainly

The consumer AI writing market is full of tools that produce text. Which of them you need depends on whether your problem is volume or stakes. Subtext is built on a more specific and more difficult premise, that the point of a personal message is to be recognisably from you, and that any assistance which erases that has taken more than it gave.

Everything follows from there. Starting from your draft. Reading emotional risk before offering a fix. Ingesting the whole thread. Asking what you want before assuming. Showing versions to choose from. Deleting the conversation once you no longer need it.

The research literature says that sounding like a machine costs you something real with the people you care about. We designed Subtext around that finding. None of those studies tested Subtext, so judge it on a draft of your own.

If you are holding a draft right now, the fastest way to understand any of this is to paste it in and watch the tags appear. Subtext runs on iOS, Android¹³ and the web at web.subtext.it, with one login across all three. Try Subtext in your browserGet the appTry Subtext in your browser

---

Think another tool does this better? Make the case on LinkedIn.

Samet Durgun is the co-founder of Subtext, an app that catches the emotional tone of your messages and rewrites them in your own voice. He’s based in Berlin.

Sources

Every link above goes to the primary source where one exists, numbered in order of appearance. Product features and store figures checked August 2026, and the Grammarly and Copilot features in September 2026.

1. Hohenstein, J., Kizilcec, R. F., DiFranzo, D., Aghajari, Z., Mieczkowski, H., Levy, K., Naaman, M., Hancock, J., & Jung, M. F. (2023). Artificial intelligence in communication impacts language and social relationships. Scientific Reports, 13, 5487.
2. Khadpe, P., Wenzel, K., Loewenstein, G., & Kaufman, G. (2025). Explaining the Reputational Risks of AI-Mediated Communication. AAAI/ACM Conference on AI, Ethics and Society.
3. Jakesch, M., French, M., Ma, X., Hancock, J. T., & Naaman, M. (2019). AI-Mediated Communication: How the Perception that Profile Text was Written by AI Affects Trustworthiness. CHI Conference on Human Factors in Computing Systems.
4. Liu, Y., et al. (2022). Perception of AI-assisted writing in condolence and support emails. CHI Conference on Human Factors in Computing Systems.
5. Fan, G., Liu, D., & Pan, L. (2026). Is It Still You? Attributing Authorship and Authenticity in AI-Assisted Romantic Communication. CHI Conference on Human Factors in Computing Systems.
6. Grammarly. Writing Tone Detector and Tone Suggestions.
7. Grammarly Support. Reader Reactions user guide.
8. Microsoft Support. Get email coaching with Copilot in Outlook.
9. Apple Support. Use Writing Tools with Apple Intelligence on iPhone.
10. Abner Li. Google Messages Magic Compose starts rolling out internationally. 9to5Google (2023).
11. Subtext. Terms, privacy and your account.
12. Apple App Store. Subtext: AI Writing Assistant, listing and user reviews.
13. Google Play. Subtext: Writing Assistant, listing and user reviews.