Machine translation is as old as AI. In 1947, Warren Weaver wrote to Norbert Wiener that, faced with an article in Russian, one could say: “This is really written in English, but it has been coded in some strange symbols. I will now proceed to decode”. The letter reappears in his 1949 memorandum (Weaver, 1949), whose cryptographic conception of translation encouraged the first projects (M. Mitchell, 2019, ch. 12), among them the 1954 Georgetown–IBM demonstration, which translated just over sixty sentences from Russian.
The history has three stages (M. Mitchell, 2019, ch. 12). First, hand-written rules: grammars of both languages plus bilingual dictionaries, brittle outside narrow cases. From the 1990s, statistical translation: probability tables between phrases, learned from parallel corpora such as the bilingual proceedings of the Canadian Parliament. And from the mid-2010s, neural translation with encoder–decoder architectures (Seq2seq, Bahdanau attention); Google Translate adopted it in 2016. The Transformer was introduced, precisely, as a translation model, evaluated on English–German and English–French (Vaswani et al., 2017).
Tiny example (from Bar-Hillel, 1960): “Little John was looking for his toy box. Finally he found it. The box was in the pen.” To translate pen one must choose between “writing pen” and “playpen”, and what decides is not grammar but knowing the relative sizes of boxes, pens and playpens. Bar-Hillel concluded that fully automatic high-quality translation was out of reach. Today the question is reversed: if a system trained only on text gets such cases right, where does it get that knowledge of the world from?
Current LLMs translate without having been trained specifically to do so; GPT-2 already did it in a rudimentary way, although its corpus contained hardly any French (Radford et al., 2019, sec. 3.7). Evaluation remains a problem: metrics such as BLEU compare the output with reference translations at the level of tokens (Alammar & Grootendorst, 2024, p. 374), and M. Mitchell (2019) warns that announcements of “human parity” rest on limited evaluations and that these systems translate without understanding the text in the human sense (2019, ch. 12).