Are they mad at me, or am I overthinking it?

Probably some of both, and the research can tell you which part is which.

Here is the shape of it. You are reasonably good at detecting that a message is negative. You are close to a coin flip at identifying which negative emotion it is. And you are far more confident about your reading than your accuracy justifies. So “something is off” is usually real information. “They are angry with me specifically” is usually a guess wearing a certainty costume.

I co-founded Subtext, an app that reads what a message is doing before you reply to it, so I have a stake in you finding this interesting. I will also be clear later about where a tool like mine should stay out of the way, because this topic has a line in it.

What the accuracy research actually found

The foundational study is Kruger, Epley, Parker and Ng¹, five experiments, and it is quoted more often than it is read.

The headline you have seen, that people understand tone in text only 56 percent of the time, comes from one study within it. That study forced a binary sarcastic-versus-serious judgment among a small student sample, and the paper’s own text describes email accuracy there as indistinguishable from chance. It is not a general accuracy rate for texting, and the 56 percent figure is a reading of a figure in the paper rather than a number in the text.

The study that matters more for anger is the third one, with 308 people conveying sarcasm, seriousness, anger or sadness. Email receivers believed they were right 89.3 percent of the time and were right 62.8 percent of the time. The paper publishes no anger-only rate, so anyone quoting one has invented it.

Two things worth taking from this. First, in the very first experiment accuracy was actually 84 percent, and the authors explicitly warn against concluding people are bad at this. The durable finding is the gap between felt clarity and real clarity, not incompetence. Second, and this is the one that stings: friends were no more accurate than strangers, and just as overconfident. Knowing someone well does not fix it.

Holtgraves² gives the most useful modern numbers. Senders wrote messages conveying one of 22 emotions without naming them. Exact free-response identification was 20.1 percent. With multiple choice, 46.3 percent. Anger specifically was identified around 38.8 to 40 percent.

But when readers got it wrong, they landed in the correct positive-or-negative valence about 85 to 90 percent of the time.

That is the whole answer in one statistic. Your sense that the message is negative is carrying real signal. Your conviction about which negative thing it is, is mostly not.

The negativity bias is real, and narrower than you have been told

The popular version says everyone defaults to the worst reading of an ambiguous text. The evidence is more specific than that, and this merge downgraded the claim from established to contested.

On the supporting side: Sillars and Zorn³ documented a negative intensification bias in workplace email, where receivers rated messages more negatively than uninvolved observers and were only weakly anchored to the message’s actual features. The bias was stronger in poor communication climates and among people lower in the hierarchy. Byron’s framework⁴ predicts positive emails read as neutral and neutral emails read as negative.

On the other side: Holtgraves found valence preserved 85 to 90 percent of the time, which is closer to accuracy than to a negative default. And a 2025 preregistered couples study⁵ found no significant overall tendency to rate partners more negatively than partners rated themselves. What it found instead was projection: people partly read their own current mood into their partner’s messages.

So the defensible version is not that ambiguity makes everyone assume the worst. It is that negative interpretation is predicted by who is reading and what the message looks like. Which brings us to the uncomfortable part.

Who reads ambiguity as hostility

These findings measure interpretation bias, meaning which reading you pick when there is no objectively right answer. They do not show that anxious people are worse at detecting real anger. That distinction matters and most coverage collapses it.

Kingsbury and Coplan⁶ built ambiguous texting vignettes and found higher social anxiety associated with more negative interpretations, across samples of 215 and 353. In the second study, messages attributed to female senders were read more negatively, especially by male readers.

Two meta-analyses agree on direction and disagree on size. Chen, Short and Kemps⁷ found g = 0.83 across 44 studies while flagging probable publication bias. A much larger 2026 analysis of 295 samples and 50,296 people found g = 0.48. Both real, and the newer larger one suggests a medium rather than large effect.

For rejection sensitivity there is direct texting evidence. Keane and Hammond⁸ found it predicted more negative readings of ambiguous teasing texts across samples of 490 and 394, with small correlations. Emoji, abbreviations and relationship context all shifted interpretation too, which again says tone does not live in the text alone.

Intolerance of uncertainty is the construct that fits this experience best, and its direct evidence is the thinnest. It is well established as a transdiagnostic trait, with a meta-analysis of 181 studies⁹ finding a moderate association with symptoms. But no convincing experiment shows that high intolerance of uncertainty causes more rereading, more last-seen checking, or more anger inference. The link to texting is a reasonable hypothesis, not a finding.

Things that are commonly said and are not established

“Rereading the message makes it worse.” No experiment has randomised people to reread versus leave it alone and measured anxiety or accuracy. Clinically plausible, empirically untested.

“Turn off read receipts and you will feel better.” Two cross-sectional studies point in opposite directions. One found more perceived stress among active read-receipt users; a much larger 2026 study of 2,992 WhatsApp users found people who had disabled receipts reported the highest stress. Self-selection cuts both ways, and this cannot be stated as demonstrated.

“Asking if they are mad backfires because the question reads as passive-aggressive.” Widely repeated, and no study anywhere supports it.

“Anxious people misread texts.” They interpret ambiguity more negatively. Whether they are less accurate about a real sender’s real state has not been shown.

What runs through all four is the same thing: the evidence is much better at describing what a message contains than at telling you what a person feels. That is also the line Subtext works inside. It reads language, not minds.

What actually helps

Separate what the message establishes from what you are adding. The text establishes its literal content and, fairly reliably, its broad valence. Everything past that is inference. Writing that line explicitly is more useful than another reread.

Ask once, in a way that gets you information. No study has randomised people to infer versus ask, so this is reasoning rather than a validated technique. But it has an obvious advantage: a reply introduces new evidence, and rereading recycles the same ambiguous cue. One clarifying question is information gathering. The fifth “are you sure you’re not annoyed” is something else, and I will come back to it. Getting that question to sound curious rather than accusing is harder than it looks, and it is the sort of one-line message Subtext is most useful on.

For anything genuinely fraught, move to voice. Schroeder, Kardas and Epley¹⁰ found voice reveals a thoughtful mind where text conceals it, and that speaking during disagreement produced better impressions and less perceived conflict than writing.

Then watch what you send. This is the part you control, and it is where a reading you are not certain about turns into a real problem. A reply written to a message you have decided is hostile carries that decision in it, and the other person answers the reply rather than your reasoning.

That is the job Subtext does here. Paste the thread or screenshot it, and it works on the language rather than guessing at motives: what the message actually establishes, which parts are ambiguous and what the plausible readings are, and then what your draft reply is about to signal to someone who may not have meant anything by it. If what you need is one neutral clarifying question rather than a paragraph, it will help you write that instead. Free to start, on your phone or in the browser.

Where this stops being a writing problem

Two boundaries worth naming, and the research is clear enough about both that I would rather say them than sell past them.

A tool can tell you what language is doing. It cannot tell you what a person is feeling, and any product that keeps confidently answering “they are not annoyed with you” to the same exchange submitted five times is not analysing anything. It is providing reassurance on demand. Reassurance seeking is well documented: a meta-analysis of 38 studies¹¹ found it associated with depression, and a 2023 study using daily assessments¹² found it rose on higher-anxiety days, with intolerance of uncertainty predicting more of it. The relief is brief and the loop tightens. That is worth knowing whether the reassurance is coming from an app or from a friend you have asked four times.

And if the checking is constant, if the rereading is eating your evenings, or if the anxiety persists whatever the person actually replies, that has stopped being a wording problem and no writing tool is the right answer. The best-evidenced route for the underlying distress is CBT targeting intolerance of uncertainty, where a review of 28 randomised trials¹³ found a large effect. If that sounds like where you are, talking to someone, a GP or a therapist, will do more than any amount of better-worded texting.

---

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

Every link above goes to the primary source where one exists, numbered in order of appearance. Where a popular claim has no study behind it, this post says so instead of repeating it. Checked 22 August 2026. This piece touches anxiety and reassurance seeking; if that is where you are, a doctor or therapist is a better resource than an article.

1. 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. Five experiments; the widely quoted 56 percent figure is a reading of Figure 1 in Study 2, not a general accuracy rate. doi.org
2. Holtgraves, T. (2022). Implicit communication of emotions via written text messages. Computers in Human Behavior Reports, 7, 100219. N = 136 and N = 167. doi.org
3. Sillars, A., & Zorn, T. E. (2021). Hypernegative Interpretation of Negatively Perceived Email at Work. Management Communication Quarterly. Sample size and funding unverified. doi.org
4. Byron, K. (2008). Carrying too heavy a load? The communication and miscommunication of emotion by email. Academy of Management Review, 33(2), 309-327. Theory paper, no sample. doi.org
5. Steinebach, P., Stein, M., & Schnell, K. (2025). Messenger-based assessment of empathic accuracy in couples’ smartphone communication. BMC Psychology. N = 102 in 51 couples; recruitment fell short of the preregistered target. doi.org
6. Kingsbury, M., & Coplan, R. J. (2016). RU mad @ me? Social anxiety and interpretation of ambiguous text messages. Computers in Human Behavior, 54, 368-379. N = 215 and N = 353. doi.org
7. Chen, J., Short, M., & Kemps, E. (2020). Interpretation bias in social anxiety: A meta-analysis. Journal of Affective Disorders. 44 studies, N = 3,859; authors flag probable publication bias. doi.org
8. Keane, S., & Hammond, M. D. (2024). Rejection sensitivity and interpretation of ambiguous texts. Canadian Journal of Behavioural Science. N = 490 and N = 394. doi.org
9. McEvoy, P. M., Hyett, M. P., Shihata, S., Price, J. E., & Strachan, L. (2019). The impact of methodological and measurement factors on transdiagnostic associations with intolerance of uncertainty. Clinical Psychology Review. 181 studies, N = 52,402. doi.org
10. Schroeder, J., Kardas, M., & Epley, N. (2017). The Humanizing Voice. Psychological Science. A claimed N of 1,576 conversation partners could not be verified against the paper. doi.org
11. 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. 38 studies, N = 6,973. doi.org
12. Meyer, A., Silva, K., & Curry, J. (2023). Reassurance seeking in daily life. Behaviour Research and Therapy. N = 105, 14-day ecological momentary assessment. doi.org
13. Miller, K., & McGuire, J. (2023). Targeting intolerance of uncertainty in treatment. Journal of Affective Disorders. 28 randomised trials, g = 0.89. doi.org