Lost (and Found) in Machine Translation

AIn 1954, journalists crowded into IBM's New York headquarters to watch a computer translate Russian sentences into English. The demonstration — sixty carefully chosen sentences, a vocabulary of 250 words — was a sensation, and the researchers' prediction made headlines: within five years, perhaps three, the problem of translation would be solved. It was a confident forecast for a field about to spend forty years failing. Early systems worked like brilliant clerks with dictionaries and grammar rules, and produced legendary nonsense; a whole generation of funding collapsed in the 1960s when an official report concluded machine translation was slower, cruder and dearer than the human kind, with no prospect of improvement.

BThe field advanced only when it abandoned understanding altogether. From the late 1980s, statistical systems learned from millions of already-translated sentences — parliamentary records from bilingual Canada and the European Union were a favourite quarry — and simply calculated which output was most probable, knowing nothing of grammar or meaning. Quality jumped. The decisive leap came in the 2010s, when neural networks trained on vastly larger collections began producing sentences that were not merely intelligible but fluent, and services offering instant translation across a hundred languages became a free, unremarkable part of daily life. A traveller today points a phone camera at a menu and reads it in their own language — casually performing what the 1954 optimists promised and could not deliver.

CYet fluency turned out to conceal failure rather than abolish it, and the failures have a pattern. Idioms translated word by word produce confident absurdity. Jokes, poetry and wordplay — where the form is the meaning — usually arrive dead. More insidiously, systems trained on human text inherit human habits: in languages that lack grammatical gender, sentences about doctors and nurses have famously returned from translation with the doctor male and the nurse female, the statistics of the past presented as fact. And because the machine always produces something, with equal smoothness whether it is right or catastrophically wrong, the errors are precisely the kind a trusting reader cannot see. A mistranslated menu is comedy; a mistranslated medical consent form or asylum interview is not — and documented cases of both exist.

DThe deeper limits are about what translation actually is. A sentence's meaning depends on speaker, situation and culture: a Japanese refusal phrased as 'that would be a little…', an English 'quite good' delivered with a pause — the words are the smallest part of the message. Human translators do not convert text; they make judgements about what the author meant to do to the reader, and remake that intention with different materials. This is why literary translators quarrel for pages over a single line of poetry, and why the profession's response to the machines has not been extinction but reorganisation: routine technical matter goes to the software, and human effort moves to editing its output and to the texts where nuance is the cargo.

EThe likeliest verdict on the 1954 prediction is that it mistook the finish line. Translation as substitution — this word for that — is largely solved, and solved so well that billions of conversations now cross language barriers that would have silenced them a generation ago; that achievement deserves more wonder than it gets. Translation as understanding remains untouched at the top: the machine has no idea what it is saying, and for most purposes it does not matter, until suddenly it does. The practical wisdom, seventy years on, fits in a sentence: trust the machine with information, and a human with intention.