n everyday language, words are constantly being created, and these words either persist or disappear. Although this phenomenon has been the subject of much linguistic research, the factors which influence the fate of a new word remain largely unknown, partly because of the difficulty of recording spontaneous language use over time. Examining the varieties of language used on social media allows us to overcome these limitations. We collected over 650 million tweets written in French, covering several years of ordinary interactions between 2.5 million users. We also collected the network of social links between these users. We identified nearly 400 words that appeared in the corpus between 2012 and 2014, and tracked their diffusion over 5 years within the network of users. Some of these words lead to changes, while others generate only ephemeral buzz. By looking at the position in the network of users who adopt these innovations, we show that words adopted by users who are more central in their community and easily in contact with other communities become established in the language, and vice versa. Thus, the position in the network of speakers who adopt these words is enough to predict their fate.