This study presents empirical evidence from read vs. interactive speech to shed light on the nature of the alveolar-retroflex sibilant merger by young speakers of Taiwan Mandarin (TM). TM speakers often merge the two sibilants through deretroflexion of the retroflex category. The results of the read speech showed that the variation is on a full continuum, from a complete merger to clear contrasts, and the merger is more prevalent among male speakers, demonstrating the impact of the social stigma associated with the merger. However, the results of the interactive task demonstrated that speakers who merged the contrast produced the retroflex sounds as distinct from their alveolar counterparts, revealing hidden structures in the mental lexicon. The mismatch between the abstract phonological knowledge and actual implementation in production suggests that the exposure to phonological systems of other speakers, especially those who make clear distinctions, has led to the incorporation of discrete categories into the phonological knowledge of the merged speakers. These findings suggest that large individual variation in the early stages of sound change may provide evidence for possible categories in a given language for language learners; however, their implementation may be further modulated by social as well as other phonetic factors.