The METEO System, built by the TAUM (Traduction Automatique a l’Universite de Montreal) group for Environment Canada, is widely cited as the first machine translation system put into genuine daily operational use. The first version went into operation in 1977 to translate the country’s weather bulletins between English and French, a legal requirement in officially bilingual Canada that had previously consumed the time of junior human translators doing repetitive work.
METEO succeeded because it tackled a deliberately narrow domain. Weather forecasts use a small, controlled vocabulary and a limited set of sentence patterns (“cloudy with a chance of showers,” “winds northwest 20 kilometers per hour”), which made high-quality rule-based translation tractable in a way that open-ended text was not. The system translated tens of thousands of words a day and, unusually for its era, produced output good enough to use without human revision for most bulletins, automatically flagging the roughly one-fifth of input it could not handle confidently so that a human could take over.
The TAUM group later attempted a far harder problem with TAUM-AVIATION, a system for translating aircraft maintenance manuals. The 1985 evaluation paper “TAUM-AVIATION: Its Technical Features and Some Experimental Results,” published in Computational Linguistics, documents both projects and the lessons learned about how much harder unrestricted technical text was than controlled weather language.
METEO remained in service for roughly two decades, a remarkable run for any software. Its lasting lesson is one that recurs throughout AI history: a system aimed at a tightly bounded, repetitive task can deliver real operational value long before general-purpose versions of the same technology become reliable.