Does translation change meaning? Translation studies says yes, and treats the change as systematic rather than accidental. The evidence is strong for some kinds of change and weak for others, particularly the claim that a single mistranslation altered history.
What is driving the interest now
I could not verify a dated news development behind the current interest, so the most concrete evidence is from 2019. A peer-reviewed NAACL 2019 paper, “Lost in Machine Translation: A Method to Reduce Meaning Loss”, argues that machine translation can erase distinctions that matter. An arXiv study, “Lost in Translation: Loss and Decay of Linguistic Richness in Machine Translation”, reports similar losses. Neither was checked against current systems.
What is new is the machine translation evidence. The underlying idea is not. The arXiv study summarises Berman (2000) as describing “deforming tendencies” in translation, including lexical impoverishment, which a translator can mitigate but largely cannot avoid. That is a secondhand summary, and I did not check the original.
Where meaning shifts between languages
Documented cases usually involve one of four mechanisms: lexical gaps, where one language has no single word; grammatical or social distinctions that one language must encode and another does not; connotation and wordplay; and cultural idioms.
Grammar and politeness: tu and vous
French separates informal “tu” from formal “vous”; English has one “you”. A popular language blog notes that translating both as “you” loses the difference. It is a weak source, so treat the example as well known rather than rigorously measured.
Idioms and connotation
Spanish “estar en las nubes” means “to be daydreaming”, though its literal sense is “to be in the clouds”. A word-for-word rendering keeps the image and drops the meaning. The example is standard, but I found no primary source for it.
Why “untranslatable” is usually the wrong word
Linguists call a missing single word a lexical gap. A gap is not an idea that cannot be carried across. An EBSCO research starter says a text is untranslatable only if something non-trivial is lost even in the best approximation. Many popular lists of untranslatable words give a definition, and a definition is itself a translation.
The idea is often tied to the Sapir-Whorf hypothesis. Its strong form, that language determines thought, is widely rejected. Tomedes, a commercial translation company’s blog, says empirical research has largely discredited it, while the weaker form, that language influences some thinking, retains significant support. The Decision Lab reaches a similar view of the strong form. Both are secondary sources, so use them as orientation.
What machine translation does to distinctions
The clearest documented case is the NAACL paper’s example. “I cut my finger” and “I cut my finger off” describe different states of the world, yet Fairseq and Google Translate both rendered them as “Je me suis coupé le doigt.”, which does not show whether the finger is detached.
The paper argues that translation systems behave as many-to-one functions, so distinctions collapse and cannot be recovered. That is a claim about how these systems work, not a measured error rate. Whether current tools still produce this output was not verified.
The mokusatsu case and the disputed causal claim
The most cited historical case concerns the 1945 Potsdam Declaration. Prime Minister Suzuki answered with “mokusatsu”, a word that Britannica’s account shows has been read as refusal or as “no comment”. In 1950, Kazuo Kawai argued that the bombings might have been avoided had the word been translated as “withhold comment”.
A Pacific Historical Review article argues that Kawai’s account cannot withstand serious scrutiny and that neither reading implies anything is held back. An online article goes further and argues there was no translation error at all. The causal claim is therefore contested. I found no authoritative summary of scholarly consensus, so treat “a mistranslation changed history” as an open question, not a fact.
Isaiah 7:14 and the limits of lexical evidence
The Hebrew “almah” in Isaiah 7:14 is a second contested case. The Septuagint renders it “parthenos” there and in Genesis 24:43. Wikipedia reports that most scholars agree the passage did not intend a miraculous conception, while “virgin” can be an appropriate translation depending on context.
A Norwegian Bible committee chose “den unge jenta”, meaning “the young girl”, which does not explicitly refer to virginity, according to a University of Arizona Bible Interpretation account. The debate is theological as well as linguistic, and confessional websites argue one side without neutrality. The lexical evidence shows the word admits more than one rendering; it does not settle what the author meant.
Loss is not the only outcome
Translation studies distinguishes inevitable loss, which comes from differences between languages, from avertable loss, which comes from translator choices. A presentation deck uses this framing, though I could not trace its original source. Gain is the counterweight: a university handout argues the field has overlooked how translators can clarify or enrich a text, citing a scholar it does not name. That corrects the record without refuting loss.
What to do with this
- Check whether a claim names a source you can open.
- Prefer translation-studies scholarship over listicles and commercial blogs.
- When you see “untranslatable”, ask whether the writer means a lexical gap. Usually they do.
- For machine translation, test the distinction that matters to you rather than trusting the output.