What Makes Translation Difficult? More Than Swapping Words

The disagreement that runs through the whole history of the practice is whether to translate the words or the sense. Cicero and Horace both wrote about it, and Saint Jerome, translating the Bible into Latin in the fourth century, said explicitly that except in scripture, where even the word order is a mystery, he rendered sense for sense rather than word for word. The two poles have carried many names since, and the terms most used now are Lawrence Venuti's: domestication, which makes a text read fluently in the target language and hides the fact of translation, and foreignisation, which preserves the strangeness and reminds the reader that this was written elsewhere. Neither is correct in general, and the choice has ethical weight, since a smoothly domesticated text is easier to read and quietly erases the difference of the culture it came from.

Certain categories of problem recur in every language pair:

• •Idioms, which are not compositional, so that a phrase about raining cats and dogs must be replaced by an unrelated phrase of similar effect or by a plain statement that loses the colour
• •Grammatical categories one language forces and another lacks, such as grammatical gender, evidentiality marking how the speaker knows something, or the distinction between formal and familiar address, which a translator into English must either drop or signal some other way
• •Register and dialect, where rendering a regional variety by another region's dialect imports a whole set of associations that were not in the original
• •Wordplay, where a pun depends on a coincidence of sound in one language and can only be replaced by a different joke in the same place
• •Culturally specific objects, foods, institutions and references, which can be kept and footnoted, substituted with a local equivalent or explained in the text, each with a cost
• •Names that carry meaning, a persistent problem in fiction where characters are called things that describe them
• •Verse, where metre, rhyme, sound and sense cannot all be preserved and the translator must decide which to sacrifice

The strong version of the linguistic relativity hypothesis, that language determines what its speakers can think, was associated with Benjamin Lee Whorf, is not supported, and rests partly on a claim about Inuit words for snow that has been comprehensively debunked. A weaker version has survived experimental testing. Speakers of languages that divide the blue range into two basic terms discriminate between those shades marginally faster, an effect that disappears when the verbal channel is occupied by another task. Speakers of languages that use absolute compass directions rather than left and right maintain a running sense of orientation far better than speakers of English, and represent time as running east to west rather than left to right. Grammatical gender influences the adjectives speakers reach for when describing an object. These are real and small, and they establish that language nudges habitual attention rather than fixing the boundaries of thought, which is enough to matter for a translator without supporting anything stronger.

Automatic translation has improved more in the last decade than in the previous fifty years. Early systems worked by rules, encoding grammar and dictionaries, and were brittle. Statistical systems from the 1990s learned from parallel texts, using large bilingual corpora including the debates of the Canadian parliament and the European Union, and were noticeably better and still clumsy. Neural systems from about 2016, and the large language models that followed, produce fluent output that handles context across sentences and is genuinely useful. The residual weaknesses are informative about what translation is: machines still struggle with pronoun reference where the source language omits pronouns, with maintaining a consistent level of formality, with terminology in specialised domains, with low-resource languages that have little parallel text, and with anything where the right rendering depends on knowing what the text is for. They also produce fluent errors, which are more dangerous than obvious ones, because there is no signal that anything has gone wrong.

Mistranslation has a documented history of consequences. Michelangelo's Moses has horns because the Hebrew word for the radiance of Moses's face was rendered as horned in the Latin Vulgate, an interpretation that shaped European art for centuries. A mistranslation of the Treaty of Waitangi in 1840 produced two texts, one in which Maori chiefs ceded sovereignty and one in which they ceded a lesser form of governance, and the discrepancy is still litigated in New Zealand. In medicine, a Spanish word meaning generally unwell was rendered as intoxicated in a Florida hospital in 1980, leading to a misdiagnosis, a delayed treatment and a young man left quadriplegic, and a settlement of seventy-one million dollars. These are the reasons that legal, medical and diplomatic translation is a regulated specialism, that important treaties designate one language as authoritative, and that professional interpreters exist in courts and hospitals in every country that can afford them.

Translation is difficult because languages carve up meaning differently, force different grammatical distinctions and carry cultural reference that has no equivalent elsewhere, so the translator must constantly choose between fluency that hides the original's strangeness and fidelity that preserves it. Idioms, wordplay, register, verse and meaningful names are the recurring problem categories. Language shapes habitual attention in small measurable ways rather than determining thought, machine translation is now fluent and fails in ways that are hard to detect, and specific mistranslations have altered art, treaties and lives.