What gets lost in translation, according to linguistics

Early on in Bangkok, I said something perfectly correct in Thai and watched it land wrong. The grammar was fine. The vocabulary was fine. But I had been a little too direct with someone I should have spoken to more gently, and I could see it on their face before I understood why.

For a long time I thought that was just me being clumsy. It turns out it has a name, and a body of research behind it. The gap I kept falling into, between saying the right words and actually being understood, is one of the most studied problems in linguistics. Here is what the field actually says about it, with the sources, because it explains why “just translate it” is never quite enough.

Most of meaning is never said

Start with a simple fact that linguists have understood for fifty years: we almost never say everything we mean.

In 1975, the philosopher H. P. Grice described what he called the Cooperative Principle: in conversation, people assume each other are being relevant, truthful, and appropriately informative, and they fill in the rest. When a friend asks “are you coming tonight?” and you answer “I have an early start tomorrow,” you never said no. But they heard no, because they assumed you were being cooperative and relevant. Grice called this an implicature, meaning that is communicated without being stated.¹

This is the first thing a word-for-word swap loses. The words carry the surface. The meaning lives in the gap between the words and the situation, and that gap is enormous. A sentence translated literally can be completely accurate and still miss the actual message, because the message was never fully in the words to begin with.

Politeness isn’t manners. It’s how language works.

Now add the part that wrecked me in that kitchen in Bangkok.

In their landmark 1987 work, the linguists Penelope Brown and Stephen Levinson argued that politeness is not decoration we add when we feel like being nice. It is a structural feature of human communication, organized around something they called face: every person’s need to be approved of (positive face) and to not be imposed on (negative face). Almost everything we say either threatens or protects someone’s face, and we constantly soften, hedge, and reshape our words to manage it.²

That softening is not optional and it is not universal in form. The exact way you show respect, defer to an elder, or make a request without sounding like a command differs from language to language. But the need to do it is everywhere. Which means a translation that gets the words right and the politeness wrong has not made a small error. It has broken the part of the message humans care about most.

The error that grammar can’t catch

Here is the concept that named my problem.

In 1983, the linguist Jenny Thomas described cross-cultural pragmatic failure: the moment when an utterance is grammatically correct but socially or contextually wrong, so the speaker’s intended meaning and the listener’s understood meaning come apart.³ Thomas distinguished two kinds. One is linguistic, using a form that does not carry the force you intended. The other is social, misjudging what is appropriate, how direct to be, how much deference to show.

What makes pragmatic failure so dangerous is that no grammar checker catches it. The sentence is correct. That is exactly why it slips through. You think you have communicated, the words came out clean, and the other person quietly registers that something was off. This is the failure mode that a focus on accuracy can never fix, because accuracy was never the thing that failed.

Respect, baked into the grammar

In some languages, you cannot even avoid the choice.

Many languages grammaticalize respect directly. The Romance languages force a pick between an informal and a formal “you,” the tú/usted, tu/vous distinction. Japanese has layered honorific registers, including keigo, the speech you use with elders, customers, and superiors. Korean shifts its verb endings by speech level. Thai ends polite sentences with particles, and softens further for elders. In all of these, you do not get to be neutral. Every sentence picks a level, and the level is part of the meaning.

So translating into these languages is not just choosing words. It is choosing a relationship: who you are to the person in front of you, and how much respect the moment is owed. Get that wrong and, as Brown and Levinson would put it, you have threatened someone’s face without meaning to.

Why a translation app struggles with exactly this

None of this is a knock on machine translation for what it is good at. For reading a menu, a sign, or a label, it is genuinely excellent. But the research is clear that the pragmatic layer, the politeness and register, is precisely where it strains.

Researchers at the University of Edinburgh showed this directly. In a 2016 paper, Sennrich, Haddow, and Birch noted that neural machine translation does not reliably control the formal/informal distinction on its own, and they had to add explicit “side constraints” to the model just to make it produce the right level of politeness.⁴ In other words, getting the honorifics right was not something the system did naturally. It had to be engineered in, one distinction at a time.

A broader 2024 review of the field reached the same conclusion: machine translation continues to struggle with contextually appropriate output, including honorifics, speech-level shifts, and other sociolinguistic norms, and a translation can be technically accurate while remaining pragmatically inappropriate.⁵ That phrase, technically accurate while pragmatically inappropriate, is the academic description of my Bangkok kitchen.

What this actually means for talking to people

Put the pieces together and a clear picture emerges. Meaning is mostly unsaid (Grice). Politeness is structural, not optional (Brown and Levinson). A grammatically perfect sentence can still fail (Thomas). Many languages force you to encode respect in every sentence (honorifics). And the part machines find hardest is exactly this pragmatic layer (Sennrich and others).

What all of that points to is a distinction the translation field itself has long recognized: there is a difference between converting words and carrying meaning. When you are reading the world, a translator is the right tool. When you are talking with a person, the thing you need is closer to what a human interpreter does: carry the intent, the tone, and the right level of respect, not just the dictionary entries.

That is the whole reason I stopped trying to translate at people and started building something that interprets for them instead. The linguistics had been telling me, the entire time, why the screen-in-the-face approach felt so hollow. The words were never the problem. The meaning around them was.

A note on the research

Honesty matters here, so two caveats. First, this is a living field, and the foundational frameworks have been critiqued. Anna Wierzbicka, among others, argued that the Grice and Brown-and-Levinson models carry an Anglo-centric bias and do not map cleanly onto every culture.⁶ That critique is part of why “the right way to be polite” genuinely differs across languages rather than following one universal script. Second, machine translation keeps improving, and the research above documents an active problem, not a permanent verdict. The point is not that machines are hopeless. It is that the hardest part of being understood is the part that lives beyond the words, and that is worth taking seriously.

If you have ever said the right thing and watched it land wrong, you were not imagining it. You were running into one of the most documented gaps in language.

RoamSpeak is an interpreter for real conversations: it carries your meaning out loud, in the right register, so two people can simply talk. It’s on the App Store.

Footnotes

1. Grice, H. P. (1975). Logic and Conversation. In P. Cole & J. Morgan (Eds.), Syntax and Semantics, Vol. 3: Speech Acts (pp. 41–58). Academic Press. ↩

2. Brown, P., & Levinson, S. C. (1987). Politeness: Some Universals in Language Usage. Cambridge University Press. Overview and critique (ERIC) ↩

3. Thomas, J. (1983). Cross-Cultural Pragmatic Failure. Applied Linguistics, 4(2), 91–112. Discussion ↩

4. Sennrich, R., Haddow, B., & Birch, A. (2016). Controlling Politeness in Neural Machine Translation via Side Constraints. Proceedings of NAACL-HLT 2016. PDF ↩

5. Overview and challenges of machine translation for contextually appropriate translations (2024). PMC. Article ↩

6. Wierzbicka, A. (1991). Cross-Cultural Pragmatics: The Semantics of Human Interaction. Mouton de Gruyter. ↩