This paper explores translation trainees’ attitudes toward MT in terms of trust and reliance. The data were collected within the LeMaTTT project, an empirical investigation of the potential impact of neural machine translation (NMT) on the development of info-mining and thematic competences in legal translator trainees. The sample consisted of MA-level trainees with different educational backgrounds. Besides completing either a post-editing or a from-scratch translation task, they replied to a pre- and a post-task questionnaire investigating whether and how they use MT for specialised translation tasks and their trust in and perception of its output. The analysis considers within- and cross-group tendencies and takes account of the differences in (a) the types of tasks and (b) the participants’ experience and training in specialised translation and post-editing. The trends concerning the use and perception of MT are also correlated with perceived task difficulty and self-assessment, as reported in the questionnaires. Results suggest that MT is perceived as a reliable tool which speeds up the translation process and provides candidate terminological equivalents, though revision is required. It appears to generally reduce the perceived difficulty of the ST while increasing the perceived quality of the TT in less experienced and competent trainees.