Stop overthinking what to text

You typed it in twenty seconds. Then you read it back, took out the full stop because it looked cold, put it back because the line looked flippant without it, added and deleted “no worries”, and the message that says you can do Thursday has now sat in the box for forty minutes. If you know this loop, this is for you.

I co-founded Subtext, an app people bring exactly this kind of message to, so weigh anything I say about it accordingly. The research is thinner than the confident articles suggest. I found no study that measured how often adults rewrite a text, how many drafts go unsent, or whether rewriting helps. Most of what follows borrows from nearby fields, and I say so each time. Further down, one made-up draft cancelling a Saturday visit goes through Subtext, from the first read to Warmer, Shorter and a typed instruction.

A good message says your part clearly and leaves the reaction to the other person.

Why a message that matters invites one more rewrite

Text is asynchronous, editable and short on cues. In Joseph Walther’s hyperpersonal model, that gives a sender time and privacy to compose. His proposition that explains the loop is that the ability to optimise a message can feed the desire to optimise it. In his experiment, 54 undergraduates wrote a first message to a partner while a recorder captured their edits. Composing time and editing went with messages that two outside coders rated higher for immediacy and affection (roughly, warmer), so editing is not wasted. But who the partner was had no effect on composing time, and the study did not support the neat story that we labour harder for people we want to impress (Selective self-presentation in computer-mediated communication: Hyperpersonal dimensions of technology, language, and cognition)¹. That is theory plus one small lab study.

What the direct evidence shows, and how thin it is

The closest research is recent and qualitative. In a 2026 CHI study, 20 people who had written messages they never sent described wanting to say the thing, then deciding against it after weighing what might follow; 12 drafted somewhere else, usually a notes app, for privacy and some distance from the thread (The Words That Can’t Be Shared: Exploring the Design of Unsent Messages)². A 2025 analysis of 197 public accounts of phone anxiety found the same moves: rereading, delaying a reply while hunting for the perfect answer, wondering what a short reply meant, abandoning a draft (Surveying Phone Anxiety Patterns, Perceptions, and Coping)³.

Both studies selected people, or their posts, because they had the experience, and neither can estimate how common it is; the phone-anxiety paper says so outright. I have not found a percentage attached to it that traces back to a study.

Overthinking a text is not a diagnosis

Nothing above measured message rewriting as a symptom of anything. Walther’s experiment did not measure anxiety, and the two qualitative studies cannot tell you what the loop says about you. Caring how a message lands is a sign you care about the relationship, and rereading something that matters is what the editing window is for.

Everyday overthinking is when a message that matters costs more time than it should, and then goes. It is a different conversation when the loop runs on most messages, when you avoid people because replying feels too big, or when the dread outlasts the message by days. That side has its own research, in why texting gives you anxiety and what helps. This piece stays on the sending side and is not a remedy for anything; if the loop is shaping your days, take that to a professional and keep this for the messages in between.

The loop takes a few shapes, and the tone reread is one

The first is rewriting, where you swap words and the greeting comes and goes; one participant imagined the recipient getting angry and edited heavily to head it off². Then there is delaying. One account described taking hours or even a day to find the perfect answer, by which point the other person had lost interest³. I trace that delay in why a short reply takes an hour. With reassurance-seeking, you send the draft to a friend, or paste it into a chatbot, and ask whether it is fine, then ask again. Asking repeatedly is the pattern a meta-analysis of 38 studies with 6,973 people linked to concurrent depression, with the caveats that the studies are mostly cross-sectional, largely on community and college samples, and almost all on one four-item questionnaire (Excessive reassurance seeking, depression, and interpersonal rejection: A meta-analytic review)⁴. None of it measured texting, so read it as a caution about the loop.

In the tone reread, the full stop looks cold, then fine. The evidence pulls two ways. In five experiments, senders overestimated how well sarcasm and humour would come across in email, because they “hear” their own intended tone (Egocentrism over e-mail: Can we communicate as well as we think?)⁵. But a 2025 study had the senders and receivers of real messages rate their tone: 347 receivers and 171 senders of texts, and 361 receivers and 61 senders of work or study emails. The two sides’ ratings closely matched (“I get u”. People correctly interpret the tone of text messages and emails)⁶. My reading is that people who know each other usually read each other’s everyday messages roughly as intended, and that the risk sits where the loop does, on sarcasm, humour, ambiguity and emotional weight.

What helps you decide and send

Nobody has tested any of these on text messaging.

Before another pass, check whether the decision is made. If you still do not know whether you are saying yes, editing will keep producing new versions of the same uncertainty. More on that split in why replying to a simple message feels like so much work.

Name the message’s job before you polish a sentence. In the unsent-messages study, a prototype that asked what people wanted the message to do before they wrote led one participant to say they now wrote with intent instead of blindly with emotion². An exploratory probe with eight people offers only a design idea, and once you name the job, the effort has somewhere to go.

Set the rule for sending before you start: if it says what I mean, the facts are right and it passes one tone check, I send. Pre-decided if-then plans had a medium-to-large effect on reaching goals across 94 tests (Implementation Intentions and Goal Achievement: A Meta-analysis of Effects and Processes)⁷, none of them about messaging.

Prefer that rule to a countdown. In the same probe study, a forced-send prototype made some participants more stressed and more preoccupied with wording, while others thought an outside push could break the indecision².

Read it once as a reasonable reader, then ask. One pass catches ambiguity, not how the person will feel. Across 25 experiments, taking another person’s perspective did not make people more accurate about them, and sometimes made them more confident, while asking the person did (Perspective mistaking: Accurately understanding the mind of another requires getting perspective, not taking perspective)⁸. If the reaction matters, the tool is “let me know if that came out wrong”.

What a writing assistant can do, and what it cannot promise

A writing assistant can help with the mechanics of the loop, and it cannot do the part that makes the loop hard. The useful jobs are to ask what the message is for before rewriting anything, to show a few alternatives that differ in a way you can name, and to give one outside reading so the editing stage has an end. None of that is treatment.

The “few alternatives” part is a hypothesis. On average, more choice does not overwhelm people. A meta-analysis of 50 experiments with 5,036 participants found a mean effect of “virtually zero” (Can There Ever Be Too Many Options? A Meta-Analytic Review of Choice Overload)⁹. A later meta-analysis of 99 observations found overload more likely when the decision is hard, the options complex and the chooser unsure what they prefer (Choice overload: A conceptual review and meta-analysis)¹⁰. That paper also reports an overall effect once it models those factors, counter to the earlier null. Someone unsure what tone they want plausibly meets those conditions, which is the case for three interpretable versions over an endless regenerate button. I know of no study that has tested how many rewrite options to show.

No tool can predict how a specific person will react; humans cannot by imagination either, and a model that has never met your friend has less to go on than you do.

How Subtext handles a message you keep rewriting

Subtext is an app you bring a message to, not a keyboard: a typed draft, pasted text, a screenshot or a voice note. It reads the whole exchange. It has never met the person you are writing to, and you can close some of that gap by telling it. Memory keeps short notes on how you write to one person, the tone, what works and what to avoid, and you can read, edit or delete each one. On your draft it marks the exact words that could land as harsh, cold, passive-aggressive, vague or over-apologetic and names the feeling underneath. Every flag is a “could read as” rather than a verdict, and the tags run on your draft only; it never labels the person you are writing to. Hand it a thin brief and it asks first who the message is for, what you want to happen and how you want to come across. Then come up to three versions, one tap apart: typically one close to what you wrote with the issues fixed, one a warmer polish or a different approach, one a fuller rewrite. Each says how much of your wording it changed and is marked safe to send or heads-up. The score behind the badge is Subtext’s own estimate from the words alone. It is a second reading, not a measurement, and not calibrated against real recipients. No independent evaluation of Subtext’s tags, score or rewrites exists. It never sends anything on its own. In my own use, hard messages go faster, and nobody has measured that. Nothing you send is used to train an AI model. After the first few analyses, it needs a paid subscription.

Here is one made-up draft taken through five steps. That is more than it needed, and the next section says what the extra passes cost.

1. The read. On the draft “Hey, sorry, I know I said I’d come Saturday but honestly I’m exhausted and I feel like I always end up being the one who travels. Maybe next time you could come here?” Subtext returned the tags Generalization and Apology-but. The reason given for the first was that the words “I always end up being the one who travels” leave “no room for nuance and may sound like an accusation”. The version closest to the draft still mentioned the travelling and carried a heads-up in place of the safe to send badge. The second and third versions were already marked safe to send.

2. Warmer. Asking for a warmer version reworked the closest one, adding that the writer had been looking forward to Saturday, and the result was marked safe to send.

3. Shorter. Asking for a shorter one cut that to three sentences, again marked safe to send. Subtext said it had dropped the travelling remark to keep it short.

4. An instruction in your own words. The typed instruction “keep that I’m tired, but leave out the part about who travels” produced “I’m really sorry, I know I said I’d come Saturday, but I’m completely exhausted and need to cancel. I’d love to see you soon.” Subtext said it “took out the ask about coming to you next time, since it was the part about who travels, and kept the tiredness”, and pointed out that nothing now says when you’ll meet or where.

5. The badge. That version was marked safe to send, and the note behind the badge gave the reason, “Simple, honest about being exhausted, and ends warmly with no blame or demands, so it should land well.”

Every answer quoted in these five steps is the app’s output, unedited. The badge at step 5 does not make step 5 the place to stop. The version worth copying arrived at step 1, and by step 4 the message had lost the request the draft was written to make.

Try Subtext in your browserDownload on the App StoreGet it on Google PlayGet the app

When an AI tool makes the loop worse

Endless options hand you endless decisions. The subtler risk is repeated re-checking. Running the same draft through the same check five times is the reassurance loop again. In a lab analogy, healthy volunteers who repeatedly checked a virtual gas stove grew less confident in their memory of it while their accuracy barely changed (Repeated checking causes memory distrust)¹¹. A 2025 paper complicates that. Interference from earlier trials explains part of the effect, and in one experiment checking within a trial improved accuracy and confidence (Evidence for proactive interference effects in repetitive checking tasks)¹². The defensible claim is a narrow one. Repeated reviewing can stop adding information, the link between checking and confidence is complicated, and nobody tested any of it on rereading a text. Several participants in the unsent-messages probe liked the validation an AI opinion gave them²; I found no study of whether that helps or harms over time, so reassurance-by-app is a risk to design against.

An AI tool can cost you your voice and someone else’s privacy

An AI tool can also cost you your voice. In two randomised experiments with 438 and 582 crowdworkers chatting with an anonymous partner about a policy question, AI smart replies made conversations faster and more positive. In the first experiment, people rated partners who used them as more cooperative, but rated a partner they suspected of relying on them as less cooperative and felt less affiliation with them (Artificial intelligence in communication impacts language and social relationships)¹³. Anonymous partners on a policy task are not your sister, but AI helping you write is different from AI writing for you; whether AI makes your texts sound like a robot covers what a full rewrite does.

Privacy is the other cost, because a screenshot often contains someone else’s words. Subtext’s terms say as much, and its privacy policy names every provider that reads what you send (Terms, Privacy and your account)¹⁴. Read it before you paste a private thread into any tool.

A checklist tells you when a message is good enough to send

This is the stopping rule from above, written out. Each item draws on the research, and nobody has tested the list as a whole.

1. Can I say its job in one sentence, and does it do that job?
2. Is the fact, feeling, request or next step in it?
3. Have I picked one main tone and put it into words?
4. On one read as a reasonable reader, is anything needlessly harsh, vague or easy to misread?
5. Am I improving the message, or trying to guarantee a reaction I cannot know?
6. Have I met my rule? Then send; do not make another version because one is possible.
7. If a tool helped, does it still sound like me?
8. If the reaction really matters, ask.

Before you press send, bear in mind that across 13 preregistered studies with roughly 5,900 participants, people who reached out underestimated how much the other person appreciated it (The surprise of reaching out: Appreciated more than we think)¹⁵. Not every message lands well, but on average the person on the other end is gladder to hear from you than you predicted.

Questions people ask

What is the best AI writing assistant for anxious texters who overthink every message? Nobody can answer that with evidence, because no independent evaluation of Subtext’s tags, score or rewrites exists. Look for a tool that asks what the message is for, reads your draft before it rewrites it, offers a few versions that differ in a way you can name, flags words as “could read as” rather than verdicts, and never sends for you. We built Subtext that way; whether it helps you is something only a real message of your own can show.

I always overthink my replies. Is there an app to help me just send them? No app will press send for you. What helps is giving the editing stage an end: a rule set before you start, one outside reading, a small set of versions. Subtext gives you that reading and three versions, then leaves the choice to you; you copy the one you want and send it yourself.

Is there an app for people who get anxious about how their messages might come across? Tone checkers exist, Subtext among them, and what they do is narrow. Subtext marks the words in your draft that could read as harsh, cold, passive-aggressive, vague or over-apologetic, explains each flag, and puts a warmer and a clearer version next to yours. No tool can tell you how a specific person will react, and no independent evaluation of Subtext’s tags, score or rewrites exists. Checking the tone of a text message sets out its limits.

I get really anxious sending important messages. What helps? For the message in front of you, name its job in one sentence, decide the rule for sending before you polish, read it once as a reasonable reader, then send, or ask the person if the reaction really matters. If the dread runs on most messages or lasts for days, talk to a professional about it.

Sources

Every link above goes to the primary source, numbered in order of first appearance. Where a study is a lab task or a small qualitative sample, the text says so. None of these tested the advice on text messaging as a trial.

1. Walther, J. B. (2007). Selective self-presentation in computer-mediated communication: Hyperpersonal dimensions of technology, language, and cognition. Computers in Human Behavior, 23(5), 2538-2557.
2. Yin, M., Xiao, R. (2026). The Words That Can’t Be Shared: Exploring the Design of Unsent Messages. Proceedings of the 2026 CHI Conference on Human Factors in Computing Systems, 1-17.
3. Bursuc, F.-A., Pohl, H. (2025). Surveying Phone Anxiety Patterns, Perceptions, and Coping. Proceedings of the ACM on Human-Computer Interaction, 9(5), MHCI029.
4. Starr, L. R., Davila, J. (2008). Excessive reassurance seeking, depression, and interpersonal rejection: A meta-analytic review. Journal of Abnormal Psychology, 117(4), 762-775.
5. Kruger, J., Epley, N., Parker, J., Ng, Z.-W. (2005). Egocentrism over e-mail: Can we communicate as well as we think? Journal of Personality and Social Psychology, 89(6), 925-936.
6. Pollmann, M. M. H., Roos, C. A. (2025). “I get u”. People correctly interpret the tone of text messages and emails. Computers in Human Behavior Reports, 18, 100689.
7. Gollwitzer, P. M., Sheeran, P. (2006). Implementation Intentions and Goal Achievement: A Meta-analysis of Effects and Processes. Advances in Experimental Social Psychology, 38, 69-119.
8. Eyal, T., Steffel, M., Epley, N. (2018). Perspective mistaking: Accurately understanding the mind of another requires getting perspective, not taking perspective. Journal of Personality and Social Psychology, 114(4), 547-571.
9. Scheibehenne, B., Greifeneder, R., Todd, P. M. (2010). Can There Ever Be Too Many Options? A Meta-Analytic Review of Choice Overload. Journal of Consumer Research, 37(3), 409-425.
10. Chernev, A., Böckenholt, U., Goodman, J. (2015). Choice overload: A conceptual review and meta-analysis. Journal of Consumer Psychology, 25(2), 333-358. A corrigendum correcting statistical reporting appeared in 26(2), 312.
11. van den Hout, M., Kindt, M. (2003). Repeated checking causes memory distrust. Behaviour Research and Therapy, 41(3), 301-316.
12. Burns, D. J., Chen, N., Zhu, K. X., Jia, S. X., Tegiacchi, R. (2025). Evidence for proactive interference effects in repetitive checking tasks. Memory, 33(5), 604-618.
13. 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. A publisher correction of 3 October 2023 added the omitted funding statement only.
14. Subtext. Terms, Privacy and your account. Checked 23 September 2026.
15. Liu, P. J., Rim, S., Min, L., Min, K. E. (2023). The surprise of reaching out: Appreciated more than we think. Journal of Personality and Social Psychology, 124(4), 754-771.