Grapheme–colour synaesthesia is a well-characterized phenomenon in which achromatic letters and/or digits automatically and systematically trigger specific colour sensations. Models of its underlying mechanisms diverge on a central question: whether triggered sensations reflect (1) an overdeveloped capacity in normal cross-modal processing (i.e., sharing characteristics with the general population), or rather (2) qualitatively deviant processing (i.e., unique to a few individuals). To test to what extent synaesthesia-like (automatic) letter–colour associations may be learned by non-synaesthetes into adulthood, implied by (1), we developed a learning paradigm that aimed to implicitly train such associations via a visual search task that employed statistical probability learning of specific letter–colour pairs. In contrast to previous synaesthesia-training studies (Cohen Kadosh, Henik, Catena, Walsh, & Fuentes, 2009; Meier & Rothen, 2009), here all participants were naïve as to the end-goal of the experiment (i.e., the formation of letter–colour associations), mimicking the learning conditions of acquired grapheme–colour synaesthesia (Hancock, 2006; Witthoft & Winawer, 2006). In two experiments, we found evidence for significant binding of colours to letters by non-synaesthetes. These newly-formed associations showed synaesthesia-like characteristics, because they correlated in strength with performance on individual synaesthetic Stroop-tasks (experiment 1), and because interference between the learned (associated) colour and the real colour during letter processing depended on their relative positions in colour space (opponent vs. non-opponent colours, experiment 2) suggesting automatic formation on a perceptual rather than conceptual level, analogous to synaesthesia. Although not evoking conscious colour percepts, these learned, synaesthesia-like associations in non-synaesthetes support that common mechanisms may underlie letter–colour associations in synaesthetes and non-synaesthetes.