Propaganda Forensics: Recovering the Generation Pipeline of an AI-Driven Influence Campaign
| Source: arXiv AI
Tags: propaganda, AI influence campaigns, Storm-1516, Llama 3, Mistral, misinformation, EMNLP
French researchers forensically analyzed the Storm-1516/CopyCop AI propaganda campaign using 2,646 articles, finding prompt instruction leaks on 50 of 84 campaign websites and evidence implicating both Llama 3 and Mistral-family models — the deepest technical analysis of a real AI influence operation to date.
Details
The PROPAGIA paper delivers the most detailed forensic analysis yet of an AI-generated propaganda campaign. Using 2,646 French propagandist articles from the Storm-1516/CopyCop campaign (disclosed by VIGINUM and INSIKT GROUP in 2025), researchers characterized the full production pipeline behind the operation. The corpus analysis reveals clear statistical fingerprints: PROPAGIA articles far exceed mainstream French press (SIPA) in vagueness, subjectivity, and negativity, while citing fewer sources — consistent with AI-generated content optimized for persuasion rather than accuracy. More concretely, 50 of 84 campaign websites leaked prompt instructions, including a verbatim ten-point editorial specification that directly explains many of these characteristics. The attribution work is the paper's most technically distinctive contribution. Rewriting-based detection aligns with INSIKT GROUP's earlier attribution to the Llama 3 family but additionally identifies Mistral-family model signatures — suggesting either a multi-model pipeline or different models used for different content categories. Appearing at EMNLP 2026's Workshop on Online Abuse and Harms, this work provides a replicable forensic methodology (topic modeling, sentiment analysis, vagueness measurement, rewriting-based detection) that can be applied to future AI influence operations.