Automatic and Manual Detection of Generated News: Case Study, Limitations and Challenges

BSTRACTIn this paper, we study the exploitation of language generation models for disinformation purposes from two viewpoints. Quantitatively, we argue that language models hardly deal with domain adaptation (i.e., the ability to generate text on topics that are not part of a training database, as typically required for news). For this purpose, we show that both simple machine learning models and manual detection can spot machine-generated news in this practically-relevant context. Qualitatively, we put forward the differences between these automatic and manual detection processes, and their potential for a constructive interaction in order to limit the impact of automatic disinformation campaigns. We also discuss the consequences of these findings for the constructive use of natural language generation to produce news items.