How do you tell if a text is sarcastic?

“Great, thanks for that.” You have been staring at it for a while, and it could be a thank-you or it could be the opposite, and the difference matters. I co-founded Subtext, and messages that could be read both ways are the largest single category of what people send us.

The research on this is more definite than you would expect. From the words alone, with no context and no markers, you mostly cannot tell, and the experiment showing that is thirty years old. What restores the difference is context, shared history, and a small set of markers that people use with surprising consistency. Knowing which of those you have in front of you is most of the answer.

The experiment that explains the problem

Bryant and Fox Tree took utterances from spontaneous speech, some originally ironic and some not, and presented them as text alone, with no audio and no context¹. Readers rated them as equally ironic. Adding the audio, or adding the context, restored the distinction.

That is the mechanism behind every other finding here. Sarcasm in speech carries acoustic cues, chiefly lower average pitch, reduced pitch variation and changes in voice quality². Text strips them out. Whatever is left has to come from somewhere else.

I’d note that the same authors later argued there is no single dedicated ironic tone of voice, and that the acoustic cues differ between languages³. So even in speech the signal is less reliable than intuition suggests.

How badly people do, with the numbers labelled

Kruger, Epley, Parker and Ng ran five experiments on tone over email and people quote the paper with figures that rarely say which study they came from⁴.

In the first, twelve students in six pairs sent serious and sarcastic statements. Senders predicted readers would decode 97 per cent correctly; readers managed 84. In the second, 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. Senders expected readers to decode about 78 per cent whichever channel they used, and readers were around 90 per cent confident they had read each line correctly, 89 per cent over email and 91 per cent by voice.

You will see the second study quoted as 78 per cent predicted against 56 achieved. The paper’s own Figure 1 prints both numbers: senders anticipated 78.0 per cent and email readers got 56.0. But those come from 29 pairs judging whether short statements were sarcastic or serious, so they are not a general rate for reading tone and I would not quote them as one.

Two things to carry. The samples are small mid-2000s undergraduate groups and nobody has directly replicated the paper at high power. And the mechanism is egocentrism. The sender hears the sarcastic tone in their head while typing and cannot subtract it. I have written about that separately in does this text sound rude.

And there is a gap I don’t think anyone has filled. No published human study gives a clean accuracy figure for written sarcasm with conversational context, comparable to the chance-level figure without it. People quote machine classifier scores in that slot and those scores measure something different, which I go into in what AI can and cannot tell about the tone of a message.

The markers people use

Thompson and Filik ran the anchor study, and the numbers are more specific than most summaries⁵.

Emoticons appeared in 88 per cent of sarcastic comments against 51 per cent of literal ones. The wink and the tongue-out face were the primary markers of sarcastic intent, reserved almost exclusively for non-literal use. The plain smile showed the opposite pattern, appearing in 81 per cent of literal praise and rarely anywhere else.

The counterintuitive result concerns the ellipsis, which was associated with criticism, negative sentiment and pausing, not with sarcasm as such. If you read three dots as sarcastic, the corpus says you are misreading them.

There was no gender effect on emoticon use or on which emoticons people chose, which refutes an earlier claim that men mark sarcasm with them while women mark emotion.

The neural evidence is the striking part. Weissman and Tanner found that an ironic wink emoji produced the same brain signatures found for verbal irony, with the size of the response tracking how strongly each participant treated the emoji as ironic⁶. The brain processes an ironic emoji like an ironic word, not as decoration.

The effect replicates across labs. One study found older readers were less likely to use the wink; a later one found emoji aided both age groups equally, with the difference attributed to whether the base sentences were ambiguous⁷⁸. So the age story depends on the materials, and I would not pick one result.

Sarcasm also does something to the message it carries. Filik and colleagues found participants rated sarcastic criticism less negative than literal criticism, and sarcastic praise less positive than literal praise⁹. It mutes in both directions, and a companion study measuring facial muscle activity found the same pattern physiologically.

This is roughly what Subtext is checking when it reads a two-way message. Whether the markers are there, whether the words support both readings, and how far apart the two readings sit. It will tell you a message is ambiguous. It will not tell you which reading the sender meant, because nothing can from the words alone, and the research above is the reason.

Why context does the work

Three theoretical camps, and they disagree. One holds that irony echoes a prior expectation and signals distance from it, so if you cannot identify what it is echoing, the ironic reading collapses¹⁰. Another holds that the ironist is pretending to be someone unwise and the audience is meant to see through it. A third weighs literal content, context, what you know about the author and any markers all in parallel, and predicts that when context is weak, markers carry disproportionate weight, which is why a single wink can flip a neutral sentence.

My read is that the empirical support for the context claim is thinner than the confident phrasing in most articles suggests. Direct experimental manipulation of shared context in written irony is scarce. What exists is consistent. In one small study, people did not perceive utterances as ironic without something prior to echo, and corpus work shows people deploy irony less and mark it more when addressing unfamiliar audiences.

In practice, sarcasm is a bet that the reader shares enough with you to see the gap between what you said and what you meant. In a thread with someone you know, that bet usually pays, which is a big part of what a text means in the first place. With a stranger, or across a cultural line, the words carry it alone and they mostly cannot.

Who reads it differently

Autistic readers, handled carefully. Group differences in irony comprehension are real in some studies, small or absent in others, and shrink substantially once researchers control for verbal ability and task format. A meta-analysis reported an overall effect of roughly g = -0.57, reduced or non-significant when researchers matched groups on core language ability¹¹. One study found near-identical low error rates between autistic and non-autistic children on structured tasks, with a gap appearing only when the task required explaining spontaneously why a speaker was sarcastic. A large share of the apparent gap is an artefact of how researchers ask the question.

Two further findings push against the deficit framing entirely. Alexithymia, difficulty identifying one’s own emotions, runs at around 50 per cent in autistic groups against 5 per cent in non-autistic ones, and it predicts emotion-recognition performance better than autism does¹². And the double empathy work found information degraded faster in mixed autistic and non-autistic chains than in either kind alone, which reframes the question from “can autistic people detect sarcasm” to “who is failing to communicate with whom”¹³.

I’d add, because articles on this assume otherwise, that for many autistic adults text is the preferred channel. A survey of 245 autistic adults found email and text ranked highly across many scenarios and phone calls ranked worst¹⁴.

Gender. The stereotype holds up and the comprehension difference does not. People rate identical comments as more sarcastic when attributed to a man, and judge sarcastic speakers more likely male. But forced-choice comprehension studies found no reliable gender difference, and neither did the emoticon study.

Age. Some decline in spontaneous sarcasm detection in older adults, with explicit markers narrowing the gap.

When it goes wrong

Byron’s framework proposes two biases in how people read emotion from text: they read positively intended messages as flat, and ambiguous ones as more negative than intended¹⁵. That is a theoretical model, not data, but its combination explains a familiar failure. A missed ironic compliment tends to land as passive aggression rather than as neutral confusion.

Repair is possible and people attempt it often. In a corpus of 3,750 Reddit interactions, repair initiations, meaning someone signalling they did not catch the intent, occurred in about 58 per cent of interactions¹⁶. Roughly 45 per cent of those initiations got no reply and went uncompleted. Face to face, repair almost always resolves. Online it is attempted and then abandoned nearly half the time. Which is a decent argument for checking before you send rather than repairing after, and it is the argument we built Subtext on.

So I’m left with the boring answer, which is to ask if you cannot tell and to answer if someone asks you.

The sarcasm tag

There are two literatures here and people merge them.

The computational one uses “/s” as a label. Khodak, Saunshi and Vodrahalli built a corpus of 1.3 million statements that way¹⁷. Then one group manually checked 4,334 tweets and found only 14.89 per cent of messages in the sarcasm-tagged category were actually sarcastic¹⁸. People forget the tag, use it inconsistently, or use it ironically. A marker is not clean ground truth.

The human one is nearly empty. People append tone indicators like /s, /j and /gen to declare intent, and the modern popularisation of those indicators is generally traced to neurodivergent online communities. There is abundant descriptive and advocacy material and very little peer-reviewed work testing whether they measurably improve comprehension or irritate readers. The debate between people who say tags prevent misreading and people who say tags destroy the point of irony is a debate, not a set of findings.

What to do with the message in front of you

Check for the markers first, because the wink and the tongue are the closest thing to a reliable signal that exists. Check how much history you have with the sender, because that is what the bet depends on. Read the three dots as pausing or criticism, not sarcasm, since that is what the corpus shows.

If it is still readable both ways, it is readable both ways, and the person who wrote it almost certainly does not know that. The evidence says they are far more confident it landed than they should be.

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. Bryant, G. A., and Fox Tree, J. E. (2002). Recognizing verbal irony in spontaneous speech. Metaphor and Symbol, 17(2), 99 to 119.
2. Cheang, H. S., and Pell, M. D. (2008). The sound of sarcasm. Speech Communication, 50(5), 366 to 381.
3. Bryant, G. A., and Fox Tree, J. E. (2005). Is there an ironic tone of voice? Language and Speech, 48(3), 257 to 277. See also Cheang and Pell (2009), Journal of the Acoustical Society of America, 126(3), on cross-linguistic differences.
4. Kruger, J., Epley, N., Parker, J., and Ng, Z.-W. (2005). Egocentrism over e-mail. Journal of Personality and Social Psychology, 89(6), 925 to 936. Study 1: six dyads, 97 against 84. Study 2: 29 dyads analysed, chance against roughly three quarters. The 78 against 56 pair is the email condition, predicted against achieved, as printed on the paper’s Figure 1. Never directly replicated at high power.
5. Thompson, D., and Filik, R. (2016). Sarcasm in Written Communication: Emoticons are Efficient Markers of Intention. Journal of Computer-Mediated Communication, 21(2), 105 to 120. Experiment 1 n = 51, Experiment 2 n = 113, Glasgow participant database, not restricted to students. Funded by ESRC grant ES/L000121/1.
6. Weissman, B., and Tanner, D. (2018). A strong wink between verbal and emoji-based irony. PLOS ONE, 13(8), e0201727. Three ERP experiments.
7. Howman, H. E., and Filik, R. (2020). The role of emoticons in sarcasm comprehension in younger and older adults. Quarterly Journal of Experimental Psychology. Eye-tracking, readers aged 18 to 30 and 65 plus.
8. Garcia, C., Turcan, A., Howman, H., and Filik, R. (2022). Emoji as a tool to aid the comprehension of written sarcasm. Computers in Human Behavior, 126, 106971.
9. Filik, R., Turcan, A., Thompson, D., Harvey, N., Davies, H., and Turner, A. (2016). Sarcasm and emoticons: Comprehension and emotional impact. Quarterly Journal of Experimental Psychology, 69(11), 2130 to 2146. Companion physiological study in Psychophysiology, PMC4999054.
10. Sperber, D., and Wilson, D. (1981). Irony and the use-mention distinction. In Cole (ed.), Radical Pragmatics. With Clark and Gerrig (1984) on pretence and Kreuz and Glucksberg (1989) on echoic reminder, all in Journal of Experimental Psychology: General, 113 and 118.
11. Kalandadze, T., and colleagues (2018). Figurative language comprehension in individuals with autism spectrum disorder: A meta-analytic review. Autism, 22(2), 99 to 117. Effect size as reported in a secondary source, not independently confirmed.
12. Kinnaird, E., Stewart, C., and Tchanturia, K. (2019). Investigating alexithymia in autism: A systematic review and meta-analysis. European Psychiatry, 55, 80 to 89. Fifteen studies; 49.93 per cent against 4.89 per cent.
13. Crompton, C. J., Ropar, D., Evans-Williams, C. V. M., Flynn, E. G., and Fletcher-Watson, S. (2020). Autistic peer-to-peer information transfer is highly effective. Autism, 24(7), 1704 to 1712. Seventy-two participants.
14. Howard, P. L., and Sedgewick, F. (2021). Anything but the phone. Autism, 25(8), 2265 to 2278. Survey of 245 autistic adults.
15. Byron, K. (2008). Carrying too heavy a load? Academy of Management Review, 33(2), 309 to 327. A theoretical framework, not an empirical study.
16. Goddard, S., and Gillespie, A. (2025). Conversational repairs on Reddit. PLOS ONE, 20(1), e0316618. Twenty-five subreddits, 3,750 interactions.
17. Khodak, M., Saunshi, N., and Vodrahalli, K. (2018). A Large Self-Annotated Corpus for Sarcasm. LREC 2018. Conference paper; labels are self-annotations.
18. Sykora, M., Elayan, S., and Jackson, T. W. (2020). A qualitative analysis of sarcasm, irony and related #hashtags on Twitter. Big Data and Society, 7(2). 4,334 tweets manually annotated.