This paper investigates neural machine translation (NMT) outputs for dislocated constructions from French into English. Dislocations are often considered to be “substandard in formal registers” (Lambrecht 1994: 12). In French, multiple copies of the subject are licit in spoken data, whereas translations into English preclude them (De Cat 2007). We analysed 436 translations of French dislocated segments in the novel Voyage au bout de la nuit (Céline 1932) and a contemporary corpus for spoken data from Corpus de Français Parlé Parisien (CFPP) (Branca-Rosoff & Lefeuvre 2016) by DeepL and Google. Beyond prototypical X, c'est dislocations, translation toolkits continue to misfire, and this might be due to the lack of spoken data in training sets of NMT.