What Subtext is built on

How your own message comes across

1. In five experiments, people who wrote emails expected readers to pick up their tone, serious or sarcastic, more often than readers actually did. Writers hear the tone they meant when they reread their own words.

   SubtextSubtext can put up to 3 tags on your own draft, each quoting the words that could come across differently than you mean.

   Kruger, Epley, Parker & Ng (2005), Journal of Personality and Social Psychology 89(6)

2. People wrote texts meant to show one of 22 feelings without naming it. Readers named the exact feeling only about one time in five, and how sure the writer felt had nothing to do with whether readers got it right.

   SubtextSubtext names the main feeling it picks up in your message, so you can check it against the feeling you meant to send.

   Holtgraves (2022), Computers in Human Behavior Reports 7

3. 200 students rated made-up text exchanges. A one-word reply ending in a full stop read as more negative than the same reply without one, and longer replies were rated more positively overall.

   SubtextIf your reply is a curt one-word answer, Subtext can tag it as a k-drop and explain why it may come across as cold.

   Kemp, Kovacic & Beyersmann (2025), Telematics and Informatics 97

Not every study agrees. With real messages between people who know each other, receivers mostly read the tone the sender meant. That is why the tags point at specific words in your draft rather than judging the whole message. Pollmann & Roos (2025), Computers in Human Behavior Reports 18

What a feeling word can and cannot tell you

1. Writers labelled the feelings behind their own texts. Outside readers, people and AI models alike, struggled to match those labels, although the AI models came closer than people in nearly every comparison.

   SubtextThe feeling tag is a best guess about how your words come across, so it always shows its reason and you decide what you meant. Subtext does not tell you how the other person feels.

   Li, Zhou, Narayanan Venkit, Islam, Arya, Wilson & Rajtmajer (2025), Proceedings of ACL 2025

Replying to what they actually said

1. People read a chat where a support agent's replies did not follow from what the customer had just said. They judged the agent and the company worse, and were less satisfied, than when the replies fit.

   SubtextWhen you add their message, the reply Subtext drafts picks up what they said instead of starting fresh.

   A scripted customer service chat, not a conversation between friends.

   Lew, Walther, Pang & Shin (2018), Journal of Computer-Mediated Communication 23(4)

2. In Reddit threads from 25 communities, someone signalled a misunderstanding in 58% of exchanges, and about 45% of those signals never got an answer.

   SubtextSubtext looks for questions of theirs you have not answered yet, and the draft answers them or asks you first.

   Public threads between strangers, not private texts.

   Goddard & Gillespie (2025), PLOS ONE 20(1)

Before and after the draft

1. In 25 experiments, trying to imagine another person's point of view did not make people better at guessing what they thought or felt. Asking them did.

   SubtextWhen your goal is unclear or the moment is a big one, Subtext asks you a few short questions before it writes, instead of guessing what you mean.

   Eyal, Steffel & Epley (2018), Journal of Personality and Social Psychology 114(4)

2. In two experiments with suggested chat replies, conversations ran faster and partners rated each other as closer. People suspected of leaning on AI replies were rated more negatively.

   SubtextYou can improve a message you wrote or have Subtext write one. When you start from your own, one version stays close to your words with only the problems fixed, and you choose what gets sent.

   Hohenstein et al. (2023), Scientific Reports 13

3. People who proofread a text by reading it out loud found more errors than people who read it silently.

   SubtextRead aloud plays any draft to you, so you hear your message before you send it.

   In the study people read aloud themselves. Listening was not tested.

   Cushing & Bodner (2022), Journal of Applied Research in Memory and Cognition 11(3)

4. Readers rated the sender of a work email as less careful, less intelligent and less trustworthy when it had spelling and grammar mistakes.

   SubtextProofread only fixes spelling, grammar and punctuation and leaves your wording alone.

   Vignovic & Thompson (2010), Journal of Applied Psychology 95(2)

What we don't claim

• No study on this page tested Subtext. Each one tested people and messages, and the Subtext line says what the app does with that finding.
• The tags, the feeling word and the safe to send score are the AI's own reading. They have not been measured against how real readers took real messages.
• Subtext is a communication assistant, not therapy. Nothing here says it treats anxiety, stops overthinking or fixes a relationship.