Errors made by human translators and machine translation (MT) systems are not only quantitatively but also qualitatively different. The reason for this, it is argued in this paper, is that human linguistic processing in general and human translational processing in particular is dialogical, whereas MT processing is monological. The human translation process is heteroglossic – i.e. it takes into account all of the different ‘voices’ inherent in the translation process and in language – and these voices are then dialogized, i.e. they are put into meaningful relations with one another in an active, back-and-forth contextualizing process. MT systems, by contrast, being based in static computing principles and Cartesian, rationalist linguistic theory, work unidirectionally and do not sufficiently contextualize structures, words and expressions. An example is provided in order to illustrate these differences between human and machine translation, and it is concluded that although MT systems can be improved by ‘dialogizing’ the way they work, it is not likely that they will ever outperform human translators