Reading passage
The dream of translating by machine is nearly as old as the computer itself. In 1949, the mathematician Warren Weaver circulated a memorandum proposing that cryptographic techniques developed during the war could crack human language just as they had cracked enemy codes. Five years later, the Georgetown-IBM demonstration translated sixty carefully selected Russian sentences into English, and newspapers confidently predicted that fully automatic translation was only a few years away. The optimism proved premature. The demonstration's vocabulary was tiny, its sentences pre-screened for ambiguity, and the system relied on a handful of handwritten rules.
The first era of machine translation was built on rules. Linguists and engineers wrote thousands of instructions mapping the grammar of one language onto another, dictionary entry by dictionary entry. The approach consumed enormous effort and produced fragile results. The apocryphal story of 'the spirit is willing, but the flesh is weak' returning from Russian as 'the vodka is good, but the meat is rotten', though probably a legend, captured a genuine weakness: rule systems could not handle ambiguity, and human language is ambiguous at every level, from words to syntax to meaning.
In 1966, the influential ALPAC report concluded that machine translation was slower, less accurate and twice as expensive as human translation, and recommended that the United States government withdraw its funding. Research budgets collapsed overnight, an episode remembered as the first AI winter for language technology. A few commercial systems survived, most notably SYSTRAN, developed for translating Cold War documents and later adopted by the European Commission. SYSTRAN proved that rules could work within narrow subject domains, where terminology is fixed and style is uniform, but it remained notoriously unreliable outside them.
The field was transformed in the late 1980s when researchers at IBM abandoned rules in favour of statistics. Their insight was radical: instead of teaching the machine grammar, let it learn from example. Given a parallel corpus, a large body of texts with human-made translations, the system could compute the probability that a phrase in one language corresponds to a phrase in another. Canada supplied the ideal training material, since its parliament had for decades published every debate in both English and French. Statistical machine translation, later commercialised by services such as Google Translate, improved steadily as corpora and computing power grew.
Yet statistical systems had a structural flaw. They translated sentences piece by piece, phrase by phrase, with little grasp of the whole. In 2016, Google replaced its statistical engine with a neural network that processed entire sentences at once, and reported overnight gains in quality approaching those of a decade of incremental progress. Neural machine translation, based on deep learning models trained on hundreds of millions of sentences, soon became the industry standard. Its most recent incarnation, the large language model, translates not by applying explicit mappings but by predicting the most plausible continuation of a text, a method that often produces startlingly fluent output.
Fluency, however, is not the same as reliability. Neural systems generate plausible language even when they do not understand the source, producing errors that are grammatically perfect and factually wrong, a phenomenon known as hallucination. Performance also collapses for the thousands of low-resource languages with little digitised text, precisely the languages where translation is most needed. In medicine and law, where a misplaced negation can be catastrophic, raw machine output remains unacceptable. Studies consistently find that professional post-editing, in which a human corrects the machine's draft, is faster than translating from scratch but still essential for high-stakes texts.
The consequences for the translation profession have been paradoxical. Machine translation has not eliminated human translators; it has redefined them. Routine, high-volume work has migrated to machines, while humans concentrate on literary texts, marketing, certified documents and the supervision of machine output. The European Union's institutions, for instance, use neural translation as a first pass for millions of pages but continue to employ thousands of linguists. Demand for translation overall has risen rather than fallen, because cheap machine translation makes multilingual communication viable in situations where no translation would previously have been commissioned at all.
Seventy years after Weaver's memorandum, machine translation illustrates both the triumph and the limits of artificial intelligence. The statistical turn taught the field that scale can substitute for explicit knowledge; the neural turn showed that fluency can emerge from prediction alone. What remains elusive is genuine understanding, the ability to know what a text means rather than how it tends to be rephrased. Whether that distinction ultimately matters for practical translation, or merely for philosophy, is a question the next generation of research will have to answer.