The aim of this monograph is twofold. First, it presents the modalities of integrating a “predictive” device into the editorial process of an “alternative” publishing house. This particular firm sought to publish continuously high volumes of short texts, partly selected by means of an “algorithm” that automated stylometry (the analysis of the syntactic, lexical and semantic components of a text) and was equipped with automatic learning capabilities. Though it proved to be a useful addition (albeit it one that could still be improved), the system was discontinued a few years after it was developed, to return to a much more traditional method of manuscript selection. The second aim of the article is thus to help understand the reasons for this reversal, which related not so much to the machine itself as to the “values” and social representations with which the various stakeholders associated it.