The aim of this study was twofold: first, to develop an experimental technique as a tool to investigate learning outcomes of spontaneous, naturalistic second language (L2) learning under controlled laboratory conditions; and second, to explore how this technique can be used to understand the basic conditions and limits of this learning. Two variants of a dialogue game were tested in which corrective L2 input was provided to the learners, but the learning aspect was camouflaged. Participants were German learners of Dutch who are known to display persistent grammatical gender errors in Dutch owing to incorrect first language (L1)-L2 transfer. In Experiment 1, the participant and a 'virtual partner' (audio-recordings) took turns in describing cards using gender-marked article-noun phrases. However, the majority of the participants became aware of learning articles as goal of the experiment, either because of the way we asked participants about the goal of the experiment or because of the task used. Therefore, we changed both aspects and used a dialogue-memory game in Experiment 2, which indeed led only a minority (28%) to suspect the real goal of the study. In both experiments, participants showed substantial learning of word gender (on average 13.8 percentage points increase in accuracy) after only one instance of correct input. A manipulation of the number of trials (lag) between correct input and production did not affect results. Furthermore, the 72% of 'naïve' participants in Experiment 2 showed as much learning as the full sample. Thus, the new paradigm offers important insights into the determinants of naturalistic learning.