Why did your text sound fine to you and wrong to them?
What a 2005 email experiment says about the tone only you can hear, and what we took from it for an AI
Our position is that the tone of your message lives in your head, not in your words. We base that on a 2005 study by Justin Kruger, Nicholas Epley, Jason Parker and Zhi-Wen Ng. Sarcasm sent by email was understood about half the time, no better than chance. The same kind of lines read aloud landed nearly three times in four. The senders were just as confident either way. The step from that finding to our position is ours, not the researchers'.
What did the email study find?
In one experiment, 30 pairs of students each picked a list of statements, half serious and half sarcastic. One person in each pair typed them into an email. The other read them aloud into a tape recorder. Then each sender guessed how many of their statements the partner would read correctly, and the partner tried.
Partners who listened to the recordings decoded nearly three quarters of the statements. Partners who read the emails were right about half the time, which the authors call indistinguishable from chance. The senders' confidence did not change with the medium. In this experiment the email senders were overconfident and the voice senders guessed about right.
A larger experiment with 154 pairs widened the test to tones such as anger and sadness, between friends as well as strangers. Senders expected 88.8% of their messages to land. 70.4% did. By email the gap was widest: 89.3% expected, 62.8% landed. Face to face it was 89.4% against 73.9%, and by voice 89.4% against 73.3%.
Why? The authors' explanation is egocentrism. When you write a sarcastic line, you hear it in your head the way you mean it. The reader does not get that soundtrack. They get the words. In a later experiment, senders who were made to read their lines aloud in the opposite tone from the one they meant stopped being overconfident. Once they could not hear their own meaning, they guessed the reader's odds correctly.
What the study leaves open
This is email in 2005, between students, in a lab. There were no emoji, no voice notes, and no twenty-year habit of texting. Readers today may be better at it, or just more used to guessing. The study does not say. The numbers above are the paper's; the figures that circulate online, such as "people are 90% sure," are not.
It also measures a sender's prediction against a reader's accuracy. It does not test whether adding a smiley or a "jk" closes the gap.
What should an AI take from this?
The study was about people, not software. Our read is that it describes a cost of text that nobody budgets for. Tone travels in the voice. Strip the voice and you keep the words and lose the meaning, and the writer is the last to notice.
People who use Eudaio talk with AIs named Daimons. You can send a voice message instead of typing, and Daimons who have a voice answer in it. We built that because a chat that is only text has the same gap this study measured, in both directions.
Our answer
Our answer is that when a text lands wrong, the words were probably fine and the tone never left your head. The study puts a number on it: email senders expected 89% of their meaning to arrive and 63% did. By voice the gap was smaller, 89% expected and 73% arrived.
If you want to hear the tone and not guess at it, a Daimon can take a voice message and answer in a voice. Create a free account and see where text runs out.
Source
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 to 936.