AI that reads the conversation before helping you reply
Subtext reads the conversation before it helps you reply, because one sentence on its own is often not enough to answer. Paste the thread, drop in a screenshot, or add the voice message they sent, and it works out who said what, summarises what the other person wants, lists what still needs an answer, and drafts a reply that picks up where they left off. This page is the evidence: why earlier messages change what a line means, what an AI can take from them, and where that stops. Subtext uses the conversation you give it. It does not understand your relationship, and we know of no tool that does. It never sends anything for you.
On iPhone, Android and in your browser. Screenshots and voice messages are never used to train an AI model.
What conversational context actually means
A message carries less than it seems. Who "she" is, what "that" refers to, whether a line answers a question or corrects one, all of it resolves against earlier turns. Grice's account of how a listener works out a speaker's meaning lists the context of the utterance and shared background knowledge among what the hearer relies on (Logic and Conversation). Conversation analysis adds that a first action makes a particular second one relevant: a question projects an answer, an offer projects acceptance or refusal, so the same word does different work depending on what it follows (Opening up Closings). People fix misunderstandings over several turns and read a correction against the trouble it repairs (The Preference for Self-Correction in the Organization of Repair in Conversation). A pronoun signals that the speaker is still on whatever the previous sentence kept in focus, on evidence from single-speaker text rather than chat (Centering: A Framework for Modeling the Local Coherence of Discourse). Context, then, is five things: the earlier turns, who said each of them, what the short words refer to, which questions are open, and the cues about how close the two people are. A single line hands over none of them.
What the evidence says about earlier turns
The measured evidence comes from benchmarks, none from people's phones, and it points two ways. In one utterance-level classification study, shuffling the order of a dialogue's turns at test time, with every target sentence intact, dropped intent scores on the MultiWOZ corpus from a weighted F1 of 96.22 to 67.91, and dialogue-act scores on DailyDialog from a macro F1 of 79.46 to 66.81 (Exploring the Role of Context in Utterance-level Emotion, Act and Intent Classification in Conversations: An Empirical Study). That is a single study, on two-party English benchmark dialogue, with an older classifier rather than a large language model. It shows that some labels depend on the surrounding turns, not how much better a drafted reply becomes. The same study found that removing one speaker's earlier lines helped on one task and hurt on another, which, in our reading, argues against pasting as much as possible. The counter-evidence is older. Recurrent and transformer reply generators were "rarely sensitive to most perturbations such as missing or reordering utterances" (Do Neural Dialog Systems Use the Conversation History Effectively? An Empirical Study), so fluent output can hide shallow use of the history. The two papers disagree, and the later one says so.
What an AI can take from a thread, and how unevenly
Three things are measurable. Knowing who said what helps. Adding speaker information improved the classification of context-dependent acts such as a bare yes or no in multi-party meeting transcripts, though not in two-person telephone calls (Who is Speaking? Speaker-Aware Multiparty Dialogue Act Classification). Position matters, and on 2023-era models accuracy was highest when the relevant material sat at the start or end of a long input and fell significantly when it sat in the middle (Lost in the Middle: How Language Models Use Long Contexts). Wording matters too. When the passage holding the answer shared few words with the question, 11 of 13 models that claim a context of at least 128K tokens fell below half their short-context score by 32K tokens, GPT-4o among the better ones at 99.3 to 69.7 percent (NoLiMa: Long-Context Evaluation Beyond Literal Matching). None of these tested chat threads, so it is an inference that the message that matters, sent three weeks ago, buried mid-paste and phrased unlike today's, is the one a model is most likely to under-weight. Subtext works from the thread you hand it, and that thread is subject to the same limits.
Why Subtext does not claim to understand your relationship
Subtext uses the conversation you give it to suggest a reply. It does not understand your relationship, for a structural reason. People design what they say around what they believe the other person knows, their common ground, which accumulates over time and which people establish differently in different media (Grounding in communication). Our own examples would be the inside joke, last night's phone call, the argument at the party, the thing agreed in another app, none of which, by construction, is in any paste or screenshot. The part a model can see is not safe either. On a benchmark of conversations where people hold different information and join and leave midway (our analogy for a group chat or partial screenshot; the paper draws none), language models "perform significantly worse than humans even with chain-of-thought reasoning or fine-tuning", and the authors do not believe current models possess an actual theory of mind (FANToM: A Benchmark for Stress-testing Machine Theory of Mind in Interactions). Those scores are from 2023 models; the field still argues over what such tests measure. Either way, no tool can predict how a specific person will react, and no independent evaluation of Subtext exists, which is why we stick to the narrow claim.
How Subtext reads the thread before it drafts a reply
Subtext takes the conversation as pasted text, a screenshot of the chat, a voice note you record, or the voice message they sent you. It reads the whole exchange, and in a screenshot it works out who is speaking from where the bubbles sit. If it cannot tell, or a pasted line could be yours or theirs, it asks. For a message you received, it summarises what the other person wants, in the language it arrived in, and lists up to five things that need an answer. Then it drafts a reply that picks up the thread and answers their points, each draft carrying a safe-to-send badge or a check-before-sending note, a second reading rather than a verdict. Hand it your own draft instead and it comes back in three versions one tap apart: one close to your wording, one warmer, one shorter and clearer. You copy the one you choose. The received-message page covers the workflow of summarising what they asked, then replying. This page is the evidence for why the thread matters and where it stops. In our own use it cuts the time a hard message takes by about 90 percent; that is the founders' anecdote, not a measurement.
The same sentence means different things in different threads
"Fine." after a long, warm message is one thing; "Fine." after "Is 7 ok?" is another, and the word is identical. "Sure" after "Are you free Friday?" accepts; "sure" after "You never listen to me" closes a door. "I have an early start tomorrow" is information, an excuse or a refusal depending on what was asked. "Yeah, that works" has no content at all without the proposal above it. "No, Tuesday" corrects something, and a reply drafted from the uncorrected version confidently answers a plan both people have already dropped. "Great job" with an eye-roll emoji is teasing in one relationship and a rebuke in another, and the difference lives in the history, not in the pixels. And "Are you still mad about Saturday?" shares no words with the message where the upset happened, which is exactly the kind of link a model is most likely to miss. These are illustrations of the theory above, not study stimuli, and nobody has measured them. They are why Subtext asks for the thread rather than the line, and why working out what a text means is harder than it looks.
How this differs from a grammar check or a one-line rewrite
Subtext reads the exchange; a grammar or tone check reads your sentence. A check on one line cannot tell what "that" refers to, which question is still open, or that the plan was corrected two messages up, because none of that is in the line. Grammarly, made by the company now called Superhuman, runs a tone detector on your own text above a minimum length and offers tone and sentence rewrites; we found no documentation of it reading a screenshot of a personal chat thread to draft a reply (checked 23 September 2026). Reply features built into a messaging app are the other model. Google's help page says Magic Compose in Google Messages uses the previous 20 messages for context (Draft messages with Magic Compose, checked 23 September 2026), a bounded window inside that app. Subtext is an app you bring a message to, so it sees whatever you paste, from any app, and nothing else. No independent evaluation compares any of these, Subtext included, so this is a description of what each one reads, not a ranking. For the single-draft job, the tone checker for text messages is the page.
What happens to the conversation you paste
Subtext receives the other person's words as well as yours, so this section says what happens to them. This is framing, not legal advice. Under the EU's data protection regulation, a person handling correspondence for purely personal reasons falls outside its scope, while the provider that supplies the means does not, and a chat can contain health, political, religious or sexual details that the regulation treats as special categories (Regulation (EU) 2016/679 (General Data Protection Regulation)). A screenshot also carries more than the thread: contact names, timestamps, sometimes a notification banner from another chat. Crop what the reply does not need before you paste it anywhere. On Subtext's side, your messages go to the AI providers that read and write them, and to Subtext's servers so you can come back to a conversation; the privacy policy names every provider. Nothing you send is used to train an AI model, and it is encrypted on the way. Delete a conversation yourself and it is gone straight away, attachments included. The optional in-app web search sends your search terms to a search provider, and you can turn it off in settings.
Where it goes wrong, and what Subtext does about it
Subtext goes wrong in five ways worth naming. The first is missing context and hidden history. Whatever you leave out, Subtext cannot use, and history never typed is beyond any tool's reach. Sarcasm is the second, and you should expect errors wherever the history alone signals it. Subtext tags your own draft only and never labels the other person's message, and telling whether a text is sarcastic is hard for people too. Then there are cropped or reordered screenshots. Cropping removes the referent, the correction and who sent which line, and the two studies above disagree on whether a model ignores or misreads reordering. With long threads, the position and wording evidence above applies, and a model is likeliest to under-weight the message that matters. The last is tone drift. In two randomised online experiments between strangers, generic reply suggestions made language more positive; a partner's actual use raised cooperation and closeness ratings, while a partner merely suspected of it was rated lower, a correlation, not a cause (Artificial intelligence in communication impacts language and social relationships). The inference we draw is that a tool that only makes you sound nicer can fail where you need firmness, so Subtext offers a shorter, clearer version and lets you edit a word yourself. When it cannot tell who is speaking, it asks.
FAQ
Is there a tool that reads the whole message thread before suggesting a reply?
Subtext does. Paste the thread, a screenshot of it, or the voice message they sent, and it reads the whole exchange, works out who said what, and summarises what the other person wants and what still needs an answer before it drafts anything. You copy the reply you choose; it never sends for you.
Is there an app that reads the whole chat thread and helps me reply properly?
Subtext reads the chat you hand it, on iPhone, Android and in the browser, and drafts a reply that picks up the thread and answers their points. Paste a draft of your own and it comes back in three versions, one tap apart, each saying how much of your wording it changed; the piece on what makes a reply land well has the research behind that. It uses the conversation you give it; what was never typed, it cannot know.
Is there an app that reads a whole chat thread and helps me reply in the right tone?
Subtext drafts from the thread; on a draft you wrote it offers three versions that differ in warmth and length: one close to your wording, one warmer, one shorter and clearer. Either way there are one-tap moves for warmer, shorter, more casual, more confident or a touch of humour. On your own draft it marks words that could read as harsh, cold, passive-aggressive, vague or over-apologetic, always as a possibility, never a verdict. There is one limit. In one pair of experiments, generic smart-reply suggestions pushed strangers' language towards the positive, as the piece on whether AI makes your texts sound like a robot explains, so check the firmer version when firm is what you need.
Is there a writing assistant that understands relationship context, not just what the words say?
No, and a tool that says it does is overclaiming. Two people build what they share partly outside the transcript, and models tested on conversations where people hold different information perform significantly worse than humans (FANToM: A Benchmark for Stress-testing Machine Theory of Mind in Interactions). Instead, Subtext reads the exchange you hand it, works out who is speaking, and asks who the message is for and what you want to happen when the brief is thin or the message is a boundary, a refusal or a confrontation. For everyday messages it reads how close you and the other person are from the tone of your own draft and keeps that in the versions.
Do I have to paste the whole conversation?
No, and more is not always better. In benchmarks, extra history hurt as well as helped a dialogue classifier, and language models used material in the middle of a long input less well. Paste the part the reply depends on: the question still open, the correction, and the message you are answering. If Subtext cannot tell whether a line is yours or theirs, it asks.
Can it tell who is speaking in a screenshot?
Subtext reads a chat screenshot the way you do, your bubbles on one side and theirs on the other. When the layout does not make it clear, it asks, because a reply drafted from the wrong side of the conversation is worse than no draft.
Does it work with voice messages?
Yes. Drop in the voice message they sent you and Subtext reads it like any other message, or record your own voice note as the draft. Both go through the same reading as text: who said what, what they want, what needs an answer.
Does the other person know I used it?
Not from Subtext. It works on your side only and never sends anything; the reply leaves from your own messaging app when you paste it there. Subtext does not label the other person either. It does not tag their message as hostile or manipulative, because it reads the words on the screen, not what the sender privately feels. The piece Are they mad at me, or am I overthinking it? covers why that gap is so wide.
Does it work in my language?
Subtext comes in 17+ languages and answers in the one you write in. The summary of a message you received comes back in the language it arrived in.
What does it cost?
Subtext gives you a few free analyses to start, then a paid subscription with weekly, monthly and yearly plans, billed through the App Store or Google Play on your phone or through Stripe on the web. You see the price in your currency before you subscribe.
Subtext works in 17+ languages and answers in the one you write in. On your phone or in your browser.
Right where you need it
Subtext is built for your phone first. Your account works on the web once you have it.
Already downloaded? Your account works at web.subtext.it, synced with your phone.
Or scan to install. Click a code to enlarge it.
Works in the browser you are already using.
- Safari
- Edge
- Chrome
- Firefox
- Opera
- Brave