Does this text sound rude?

You have read it back six times and it sounds fine. It also sounded fine the first time, which is the problem, because reading it back straight away is a test your message will almost always pass. I co-founded Subtext because this question, asked of a message the person has already reread, is the most common thing anyone brings to it.

My answer is that you cannot tell from the inside, and the reason is well characterised. You silently hear your own intended tone while you write, and you cannot subtract it when predicting how someone else will read the words without it. But the popular version overstates the damage. You are bad at conveying irony, teasing, and the difference between firm and cold. You are not demonstrably bad at conveying ordinary warmth to people who already know you. Both halves matter, and so does knowing which half your message falls into.

The study behind everything, with the numbers labelled

Kruger, Epley, Parker and Ng ran five experiments on this and people cite the paper everywhere, usually with figures that do not say which study they came from¹. I’ve labelled them here, because the unlabelled versions have caused real confusion.

In the first study, twelve students in six pairs wrote serious and sarcastic statements and emailed them. Senders predicted readers would decode 97 per cent correctly. Actual accuracy was 84. Reliable overconfidence on a tiny sample, and note that actual accuracy was still high. The finding is “worse than they thought”, not “bad”.

In the third and largest study, Kruger’s team tested 154 pairs on sarcasm, seriousness, anger and sadness. Senders predicted 88.8 per cent accuracy and readers achieved 70.4. Friends were no better than strangers. And face-to-face was no better than voice-only, which points at missing intonation, not missing gesture, as the operative loss.

In the second study, email readers detected sarcasm at a rate the authors describe as indistinguishable from chance, while people hearing their partner’s lines read aloud got roughly three quarters right. You will see that study quoted as 78 per cent predicted against 56 achieved. The paper’s Figure 1 prints both numbers, but they come from this one study of sarcastic and serious statements, with 29 pairs analysed, so they are not a general rate for reading tone.

The mechanism test is the fourth study, and it is the one that justifies everything practical in this article. Participants read their own statements aloud either in the intended tone or deliberately in the wrong one. Reading them the wrong way erased the overconfidence entirely. Your own inner voice is the culprit, and this is the experiment that shows it.

Two things to carry. No one has directly replicated the paper at high power. It fell outside the sampling frame of the big replication projects and appears in none of them. It is famous, theoretically coherent and internally consistent, and it rests on small mid-2000s undergraduate samples. And the mechanism, egocentric anchoring, has theoretical parents in the curse of knowledge and the illusion of transparency, neither of which has a gold-standard direct replication either.

This table sets out three of the five Kruger, Epley, Parker and Ng (2005) experiments, study by study¹.

Study 2’s 78 and 56 are the figures you will see quoted, and they are not a general rate for reading tone. In that study the authors describe email readers’ sarcasm detection as indistinguishable from chance, while people hearing the lines read aloud got roughly three quarters right. The paper rests on small mid-2000s undergraduate samples, and no one has directly replicated it at high power.

The finding that narrows the thesis

Pollmann and Roos tested the popular assumption directly, using real messages and asking both people². Receivers rated the warmth of a message they had actually received; the sender then rated what they had intended. Across 171 matched pairs for text messages and 61 for work emails, the ratings aligned closely, and six moderators including message length and emoji changed nothing.

Read the scope carefully, because it’s what makes the two findings compatible.

My synthesis is that people are good at conveying whether they feel good or bad, and bad at conveying irony, teasing, and the line between firm and cold, especially with people who do not know them well.

Pollmann and Roos measured valence, a single dimension. “Was this warm or cold” is a different question from “did you catch that I was joking” or “did this read as passive-aggressive”, and their null result does not licence a general claim that people understand written tone. The reader’s side of the same research is in what a text means when you are the one receiving it. It is also a single study in a companion open-access journal, and no one has independently replicated it yet.

So when you ask whether your message sounds rude, the first thing to check is which kind of message it is. If it is plain and warm and going to someone who knows you, it is probably fine. If it is dry, ironic or firm and going to someone who does not, that is where the gap lives, and it is the only kind of message Subtext bothers flagging.

Why your intent feels audible

Most people hear a voice when they read. In a general-population survey of 570 people, 80.7 per cent reported sometimes or always hearing an inner voice during silent reading, and most described it as having gender, accent, pitch and emotional tone³. The 19.3 per cent who reported no inner reading voice is a useful complication, since it means this mechanism is probably not universal.

Linguists call the related phenomenon implicit prosody: silent reading projects a default melody onto text, and that melody influences how you parse it⁴. Brain measures back it up, with commas and implicit phrase boundaries producing responses that resemble the response to spoken boundaries. It is part of why people ask whether a full stop makes a text sound angry.

The limit of this evidence is that I found no single study that measures inner-voice prosody and then links it to a specific misjudgement of your own written tone. What exists is two well-supported literatures meeting at a plausible junction, plus Kruger’s fourth study, which manipulates the sender’s vocalisation and gets the predicted result.

This is also the precise gap Subtext sits in. It reads the message with no inner voice at all, which is the one thing you cannot do to your own draft, and it reports which readings the words support rather than the one you meant.

Why rereading straight away misses things, and typos survive

The same mechanism explains the typo you missed five times.

Daneman and Stainton had people write an essay and then proofread it after either twenty minutes or two weeks, alongside a familiar and an unfamiliar essay by someone else⁵. They caught fewer errors in their own writing than in unfamiliar writing. And the two-week delay reduced the disadvantage, which pins the problem on extreme familiarity and not on authorship as such.

One correction I’d make, because the wrong explanation circulates widely. Burgoyne and colleagues tested directly whether the generation effect, meaning better memory for things you produced yourself, is why you miss your own typos, and they did not confirm it. Two eye-tracking experiments found no self-generation effect in proofreading⁶. The account that holds up is familiarity and predictability. You know what the text is meant to say, so you see that rather than what is there.

Your expertise with your own message actively hurts. That is why the copy editor with no idea what you meant does better.

What actually helps, ranked

Read it aloud in the wrong tone. Kruger’s fourth study is the evidence¹. Deliberately not performing your intended tone dissolves the overconfidence. Reading it aloud in the tone you meant may do little, because that reinforces your own phenomenology. Read it flat, or read it as if you were annoyed, and see what survives.

Wait. The easiest fix to try. Daneman and Stainton found that a two week delay helped people catch errors in their own writing⁵. I know of no test of an overnight wait for tone, but the familiarity problem is the same, and overnight is usually possible for a text.

Read it aloud for errors. Cushing and Bodner’s experiments now support this directly. They compared proofreading aloud, silently, and in a deliberately hard-to-read font⁷. Aloud improved error detection. The disfluent font either impaired it or did nothing. Participants did not predict the aloud benefit, so this is useful advice people will not adopt on intuition. I couldn’t verify the sample sizes for both experiments from the accessible materials, so no numbers.

Get an outside reader. The proofreading studies support this well for clarity and errors, since a fresh reader lacks your privileged knowledge. One caveat for tone specifically is that any negativity bias belongs to the actual recipient, so a neutral third party may not reproduce the reaction of one specific anxious or subordinate reader.

Change the format. Pilotti and colleagues found that changing the typescript restores error salience⁸. This is not the same as using a hard-to-read font, which the Cushing and Bodner work found unhelpful.

Say the thing explicitly. Giurge and Bohns tested this directly, and it addresses the best-evidenced asymmetry in this literature, which is about urgency, not rudeness. They ran eight preregistered experiments with 4,004 working adults and found receivers overestimate how fast senders expect replies to non-urgent out-of-hours email, and pay for it in stress⁹. A brief note from the sender saying there is no rush reduced the bias. Your convenient Sunday-night email reads as a demand, and the fix is four words long.

Text-to-speech. Principled, but not proven. It supplies an external voice with none of your intended prosody, which is what the fourth Kruger study suggests should help. I found no direct experiment isolating it for tone.

What Subtext does with this

Subtext is the outside reader, minus the inner voice, plus the specific check for the messages where the gap lives. It flags irony that will not travel, since figuring out whether a message is sarcastic from the words alone has its own research and its own limits, a firm sentence that reads cold, and a warm message that you have edited into a flat one. It does not flag plain warm messages to people who know you, because the evidence says those mostly land.

I cover the same gap from the product side in what AI can and can’t tell about the tone of your message, if you want to see what that reading looks like on a real message. Subtext is the outside reader without an inner voice. Try Subtext in your browserGet the appTry Subtext in your browser

Sources

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

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 a University of Illinois Board of Trustees research grant and NSF grant SES-0241544. Study 1: 12 students, six dyads, 97 against 84. Study 3: 154 pairs, 88.8 against 70.4. Study 4: 54 students. Never directly replicated at high power.
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. Measures valence only. Companion open-access title; not yet independently replicated. Funding unverified.
3. Vilhauer, R. P. (2017). Characteristics of inner reading voices. Scandinavian Journal of Psychology, 58(4), 269 to 274. General-population survey, n = 570. Self-report about reading in general. Funding unverified.
4. Alderson-Day, B., and Fernyhough, C. (2015). Inner speech: Development, cognitive functions, phenomenology, and neurobiology. Psychological Bulletin, 141(5), 931 to 965. The major review of inner speech.
5. Daneman, M., and Stainton, M. (1993). The generation effect in reading and proofreading: Is it easier or harder to detect errors in one’s own writing? Reading and Writing, 5(3), 297 to 313.
6. Burgoyne, A. P., Saba-Sadiya, S., Harris, L. J., Becker, M. W., Brascamp, J. W., and Hambrick, D. Z. (2023). Revisiting the self-generation effect in proofreading. Psychological Research, 87, 800 to 815. Two eye-tracking experiments; no self-generation effect found.
7. Cushing, C., and Bodner, G. E. (2022). Reading aloud improves proofreading (but using Sans Forgetica font does not). Journal of Applied Research in Memory and Cognition, 11(3), 427 to 436. Data at OSF. Sample sizes not verifiable from accessible materials. Funded by Flinders University seed and thesis awards.
8. Pilotti, M., and colleagues (2004, 2009, 2012). Series on familiarity and proofreading, in Journal of General Psychology, Reading and Writing, and Journal of Research in Reading. Changing the typescript restores error salience.
9. 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 experiments, 4,004 working adults. Funding unverified.