In this article the author argues that the disambiguation techniques employed in mainstream machine translation (MT), which draw, in essence, either on abstract facts about language or on organizations of knowledge about situations and events, supported by a plethora of rules for exploiting such facts or knowledge, have reached a deadlock and that, if the quantum leap to quality output is to be achieved, the computer needs to have access to concrete examples of language in use. The ultimate basis for disambiguating discourse by machine for translation purposes, it is argued, is the process of understanding carried out by humans, albeit, for reasons that are subsequently made clear, with the emphasis firmly placed on successful modelling of the text-internal mechanisms involved therein rather than the interference-making procedures favoured by the artificial intelligence approach to MT.