Implicit grammar learning with semi-artificial languages has shown that it is possible to learn novel grammatical structures implicitly and that this can lead to both explicit and implicit knowledge. Can these results be extended to natural L2 learning and to learners who already have some knowledge of the L2? In this experiment, we investigated the effect of implicit training on the processing of morphosyntactic violations whose saliency and similarity between the participants’ L1 (French) and L2 (English) varied. The experiment followed a pre-test / training / post-test paradigm. During the pre- and post-test, ERP responses were obtained from 16 participants while they judged the semantic acceptability of 432 sentences. 192 critical polar questions contained violations of past tense morphology with DID and HAD. The rest of the sentences were fillers, 120 of which contained a semantic violations. Participants also completed a short timed Grammaticality Judgment Task with confidence ratings and source attributions (Rebuschat et al. 2015). During 3 one-hour training sessions, participants heard correct polar questions and were asked to select the right answer among propositions containing semantic and critical tense violations In the first session, violations with DID elicited a positive peak in the 300-500ms and 500-900 ms windows. A broad negativity resembling an N400 followed violations with HAD. These effects disappeared in the post-test. Behavioural measures revealed that participants were more accurate but slower in the HAD condition. Their performance did not improve following training but their response time slowed down. Confidence ratings showed that performance was better than chance even at a medium-low level of confidence (2 on a scale of 4). Participants reported using mostly intuition but their performance was significantly different from chance only when using rule knowledge. In the semantic acceptability task, participants performed better in the post-test; but they rejected semantically correct sentences containing a morphosyntactic violation with HAD more than with DID. A debriefing questionnaire showed that only 5 participants noticed the critical morphological errors, but only two of them were able to provide explicit rules for the relevant structure. These results show that participants are more sensitive to the L1-like violation (HAD condition) despite the superior saliency of the DID violation. It seems that learners rely on different processes for the two violations – an attention-related response with DID, probably triggered by the phonological saliency of the violation, and actual morphosyntactic processing with HAD. Though no significant effect of session was found on accuracy, the slow-down in response time and the disappearance of the positive effect suggest that participants were starting to rely less on saliency to process violations with DID. The training may have been too short for effects to appear. Although participants show some degree of implicit knowledge, they still relied on explicit knowledge successfully in the GJT. This might be due to the explicit nature of the rest of their English instruction and therefore to a habit of relying on rules. Rebuschat, P. et al., 2015. Triangulating measures of awareness. Studies in Second Language Acquisition, 37(Special Issue 02), pp.299–334.