Measuring Vagueness-Based Persuasion Techniques in an AI-Generated Influence Campaign
Résumé
Recent advances in text generation have enabled the massive production of news-like content at scale. While these technologies are not inherently manipulative, they can be employed for efficient reformulation and transformation of existing content. This massive production capability is illustrated by the recent STORM-1516/CopyCop influence campaign, documented by VIGINUM and by Recorded Future. Both organizations describe an ecosystem disseminating multilingual press-like content. In the French case, press-like websites have circulated reworked narratives, sometimes via the impersonation of major outlets, such as France Télévisions, France Médias Monde, as well as national and regional daily newspapers including Le Monde, Le Parisien, and Ouest-France. In this paper, we introduce PROPAGIA, a corpus of French press-like articles attributed to STORM-1516, and we compare it with an in-house reference corpus of human-written articles from SIPA Ouest-France, one of the leading daily news outlets in France and an independent press group.4 Influence in news is known to rely on rhetorical persuasion techniques, in particular intentional vagueness and on overt or covert subjectivity. Large scale influence, particularly propaganda, can leverage text generation to amplify these effects. To characterize persuasive stance in PROPAGIA, we therefore assess, under controlled topic conditions, the prevalence of intentional vagueness and subjectivity relative to SIPA, using the VAGO tool deployed in previous studies. One goal of our comparative methodology, which puts side by side a corpus of opaquely sourced propaganda with a corpus of transparently sourced press, is also to provide a first benchmark for the construction of a pipeline of written press analysis within the newly established working group TRUSTEDNEWS.
| Origine | Fichiers produits par l'(les) auteur(s) |
|---|---|
| Licence |