What does this text mean?

You have the message open and you have read it four times, and each read gives you a slightly different answer. I co-founded Subtext, which reads messages for tone, so this is the question I get asked most and it is also the one I am most careful about answering.

The honest answer has two halves that most articles never put together. The person who wrote it is worse at conveying tone than they think, which is well established. And you, reading it, are probably getting the basic direction right, which is also established and much less well known. Where things break is narrower than the internet suggests, and knowing where the breaks happen is more useful than any decoding chart, which is the thing Subtext tries to hand you instead of a chart.

The sender knows less than they think

Kruger, Epley, Parker and Ng ran five experiments on how well tone travels in writing, and the result is the anchor for this whole field¹.

The one number worth knowing, the gap Subtext tries to close, comes from their third study, with 154 pairs.

The gap Senders predicted their tone would be read correctly 88.8 per cent of the time. Readers actually managed 70.4 per cent.

In an earlier study the gap was wider, with senders predicting 97 per cent and readers achieving 84.

The explanation the authors give is egocentrism. When you write a message you hear it in your own head, in the tone you meant, and you cannot switch that off to check what it sounds like without the voice. Their own summary is that people believe they can communicate over email more effectively than they actually can.

Note what this does and does not show. It shows ambiguity plus sender overconfidence. It does not show that readers skew negative. That is a different claim and it needs different evidence, which is exactly why Subtext does not start from the assumption that the worst reading is the likely one.

You are probably reading it right

This is the finding that reframes the question, and it came out in 2025.

Pollmann and Roos did something nobody had done properly: they asked both people². Participants uploaded a real message thread where they had asked a friend to do something, and rated how positive the reply felt. Then the friend who wrote it was asked what they had meant.

Across 347 receivers with 171 matched senders, intended valence averaged 7.39 on a nine-point scale and perceived valence averaged 7.25. The difference was not significant, and an equivalence test ruled out any effect larger than half a point in either direction. They repeated it with work and school emails, 361 receivers and 61 matched senders, and got the same result.

They also tested six things people assume matter: message length, emoji, the reader’s gender, the reader’s age, how close the two people were, and the reader’s neuroticism. None of them predicted misreading.

There is a caveat neither the study’s coverage nor most summaries mention, and it matters. Mean-level agreement was excellent, but the correlation between individual senders and receivers was moderate, at .33 and .38. So on average people are not reading messages more negatively than intended, and individual readings still vary a good deal around what the sender meant. Both things are true at once.

The limits the authors state: the messages sampled were mostly positive, the topics were benign, and everyone already knew each other.

That last limit is where the whole answer lives. This is also why Subtext works from a thread rather than a single message. A reply read against the twenty before it is a different object from a reply read on its own, and the research says the relationship is doing more work than any feature of the words.

So where do the misreads actually happen

Four places, and none of them is “all texts, all the time”. This is the checklist Subtext runs before it tells you how much to trust a reading.

When the reader has context the message does not carry. Sillars and Zorn asked people to supply an email they had found upsetting, then had independent observers rate the same email³. Receivers rated them more negative than observers did. The authors explain it through missing context: many of the emails had nothing overtly negative in them, and the receiver read them against a situation the observers knew nothing about. Their example is a message requiring staff to work two Sundays a month, which lands very differently on someone who goes to church.

Humour and sarcasm. This is the one place where readers genuinely fall apart. In the Kruger experiments, readers detected sarcasm in email at a rate indistinguishable from chance, while listeners hearing the same lines aloud got roughly three quarters of them right¹.

When the message is genuinely ambiguous. Kingsbury and Coplan built vignettes deliberately constructed to be readable both ways, things like “I heard about last night”, and found higher social anxiety predicted the negative reading⁴. That is a moderator finding, not a universal one, and it is worth separating from the popular version. In Pollmann and Roos’s real messages, most of which were not ambiguous, no such bias appeared. If this specific worry, whether one particular person is upset with you, is what brought you here, I have written about it in more depth in are they mad at me, or am I overthinking it.

When you are close to the person. This is counterintuitive and well documented. Savitsky and colleagues found people are more egocentric with friends than with strangers, communicate no better and sometimes worse, and feel considerably more confident about it⁵.

There is even a finding in the opposite direction to the one everyone assumes. Holtgraves found receivers read face-threatening messages as more positive than the sender intended, because senders overcorrected while softening the blow.

Those four zones, missing context, humour, genuine ambiguity and closeness, are the full list.

What you are actually doing when you decide what it means

Steinebach, Stein and Schnell built a messenger app that prompted couples to rate their own feelings and their partner’s feelings during real conversations, and collected 960 jointly rated interactions from 51 couples over ten days⁶.

They found real tracking accuracy: people did pick up something about their partner’s actual state. They also found a large assumed-similarity bias. And the comparison is the finding: in both their models, the influence of tracking accuracy was substantially lower than the influence of assumed similarity.

In plain terms, when you decide what your partner’s message means, your own current mood predicts your answer more strongly than their actual mood does.

Two honest flags. The study is underpowered by its own account, having planned 200 couples and recruited 51, with 57 per cent dropout. And the sample is culturally narrow. But it is measured in the medium this article is about, with real messages, which almost nothing else here can claim.

There is one more thing they found that runs against instinct. More messaging experience went with more projection, not less. Being a heavy texter does not make you a better reader. This is the case for having something outside your own head look at the message, which is what Subtext is, and it is a narrower claim than the one most tools in this category make.

The single thing that works

Eyal, Steffel and Epley ran 25 experiments on this and the answer is unusually blunt⁷.

Across the first fifteen, with 1,476 participants, instructing people to take another person’s perspective increased their effort and reduced their egocentrism, and did not increase accurate insight. It made them less accurate, at around d = -0.26, while sometimes making them more confident.

What reliably worked was asking the other person.

That is the whole practical section. Imagining your way into someone’s head is the move everyone reaches for and it does not work. Asking does, and the reason it feels expensive is a face problem rather than an accuracy problem: asking implies either that they were unclear or that you were not listening.

Two things make asking cheaper. Giurge and Bohns showed across eight preregistered studies with 4,004 working adults that receivers systematically overestimate how fast senders expect replies to non-urgent out-of-hours emails, and that a single line from the sender saying there is no rush removed the bias⁸. Stating intent beats leaving it to be inferred, in both directions. And other-initiated repair turns out to be a human universal: across twelve languages from eight families, every one has a shared system for signalling you did not catch something, with an interjection like “huh” appearing in near-identical form⁹.

Asking what someone meant is not a social failure. It is the most standard move in conversation there is. Subtext cannot ask the question for you. What it can do is show you which reading you have quietly settled on, which is usually the thing that makes asking instead feel less necessary than it actually is.

The decoding rules with nothing behind them

Worth going through, because these are what most search results for this question will give you.

Specific delay windows mean specific things. No evidence. Reply latency is power-law distributed and driven by interruptions, task switching and notification settings¹⁰. Delay does carry information, but relative to what that specific person normally does rather than against a clock. Two hours from someone who always answers in five minutes is a different signal from two hours from someone who answers twice a day.

“K” versus “OK” versus “Okay”. No direct study exists on these spellings. The nearest real finding is about a full stop on a one-word reply, which I have written about in what a full stop actually signals.

Message length ratios. Directly tested and null. Pollmann and Roos put message length into their regression and it came out at .05, not significant².

Specific emoji mean specific things. Contradicted. Miller and colleagues found that when people rated the same emoji rendering, they disagreed on whether the sentiment was positive, neutral or negative a quarter of the time, and disagreement rose when the rendering differed across platforms¹¹.

The 7-38-55 rule. This one deserves naming. The claim that communication is 7 per cent words, 38 per cent tone and 55 per cent body language, and therefore that text carries almost nothing, is a misapplication of Mehrabian’s 1967 work. He measured situations where a single spoken word conflicted with a facial expression about feelings, and he said himself that the ratio does not apply to written text or to messages that are not in conflict.

Leaving a message on read for a set number of hours. No evidence for any numeric claim. Read receipts mark display rather than reading, and certainly not intent.

None of these rules run inside Subtext either, for the same reason none of them survive contact with the evidence above.

What to do with the message in front of you

Start from the base rate, which is that you are probably reading it about right and the sender is more worried about being misread than you are about misreading. Then check whether you are in one of the four zones where things break: are you supplying context they did not have, is this humour, is the message genuinely readable both ways, and how close are you to this person.

If you are still stuck, the research says ask, and the research says asking is cheaper than it feels.

Subtext will show you the readings a message supports and how confident it is in each one, which is usually less confident than the reading you have already settled on. It will not tell you what someone privately feels, because nothing can do that from words on a screen, and any tool claiming otherwise is selling you the thing this article is about.

Free to start, on your phone or in the browser.

---

Think I have read a study wrong, or know research on this I have missed? Tell me 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

Numbered in the order they appear above. Where a figure could not be verified against the primary source, the text says so. Checked 28 August 2026.

1. Kruger, J., Epley, N., Parker, J., and 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 to 936. Funded by University of Illinois Board of Trustees Grant 1-2-69853 and NSF Grant SES-0241544. The 88.8 and 70.4 figures are Study 3, with 154 pairs; the 97 and 84 figures are Study 1, with six dyads. https://web-docs.stern.nyu.edu/pa/kruger_email_ego.pdf
2. Pollmann, M. M. H., and Roos, C. A. (2025). “I get u”. People correctly interpret the tone of text messages and emails. Computers in Human Behavior Reports, 18, article 100689. A companion open-access title rather than the parent journal; recent and not yet replicated. Study 2 preregistered. https://doi.org/10.1016/j.chbr.2025.100689
3. Sillars, A., and Zorn, T. E. (2021). Hypernegative Interpretation of Negatively Perceived Email at Work. Management Communication Quarterly, 35(2), 171 to 200. https://journals.sagepub.com/doi/10.1177/0893318920979828
4. Kingsbury, M., and Coplan, R. J. (2016). RU mad @ me? Social anxiety and interpretation of ambiguous text messages. Computers in Human Behavior, 54, 368 to 379. https://doi.org/10.1016/j.chb.2015.08.032
5. Savitsky, K., Keysar, B., Epley, N., Carter, T., and Swanson, A. (2011). The closeness-communication bias: Increased egocentrism among friends versus strangers. Journal of Experimental Social Psychology, 47(2), 269 to 273. https://www.sciencedirect.com/science/article/abs/pii/S0022103110002374
6. Steinebach, P., Stein, M., and Schnell, K. (2025). Messenger-based assessment of empathic accuracy in couples’ smartphone communication. BMC Psychology, 13, article 147. Preregistered. Underpowered by the authors’ own account: 51 couples against a planned 200. https://doi.org/10.1186/s40359-025-02483-9
7. Eyal, T., Steffel, M., and 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 to 571. https://psycnet.apa.org/record/2018-12981-001
8. Giurge, L. M., and Bohns, V. K. (2021). “You don’t need to answer right away!” Receivers overestimate how quickly senders expect responses to non-urgent work emails. Organizational Behavior and Human Decision Processes, 167, 114 to 128. Eight preregistered studies, 4,004 working adults. https://www.sciencedirect.com/science/article/pii/S0749597821000480
9. Dingemanse, M., Torreira, F., and Enfield, N. J. (2013). Is “Huh?” a universal word? PLOS ONE, 8(11), e78273. https://doi.org/10.1371/journal.pone.0078273
10. Kalman, Y. M., Ravid, G., Raban, D. R., and Rafaeli, S. (2006). Pauses and Response Latencies: A Chronemic Analysis of Asynchronous CMC. Journal of Computer-Mediated Communication, 12(1), 1 to 23. https://academic.oup.com/jcmc/article/12/1/1/4582956
11. Miller, H., Thebault-Spieker, J., Chang, S., Johnson, I., Terveen, L., and Hecht, B. (2016). “Blissfully happy” or “ready to fight”: Varying interpretations of emoji. Proceedings of ICWSM 2016, 259 to 268. https://ojs.aaai.org/index.php/ICWSM/article/view/14757