This paper deals with the links between the development of stochastic methods in Natural Language Processing (NLP) and the attempts of refounding a linguistics based on corpora. First, we define robust NLP, in opposition to “theoretical” Computational Linguistics, by specifying its criteria and methods. Stochastic methods can be defined as probabilistic methods using statistics on corpora. Second, we note that the users of these methods tend to confuse NLP models and models of language representation, in other words they tend to confuse NLP and research in linguistics. Finally, we show that this stand is close to “instrumentalism”, to which Popper was opposed.