Researchers reverse-engineered the machine behind a propaganda campaign, and found the prompts it ran on.
A new forensic study examines PROPAGIA, a corpus of 2,646 French-language propaganda articles from the Storm-1516/CopyCop campaign disclosed by VIGINUM and INSIKT GROUP in 2025. Comparing it against SIPA, a matched corpus of human-written French mainstream press, the researchers used topic modeling plus vagueness and sentiment analysis to isolate the telltale signs of propaganda: PROPAGIA articles were far vaguer, more subjective, more negative, and cited fewer sources than the real press. Digging into the sites themselves, the team found prompt instruction leaks on 50 of 84 PROPAGIA websites - including a verbatim ten-point editorial specification that explains several of the stylistic quirks - along with heavy repetition across articles. A separate rewriting-based detection method backs up INSIKT GROUP's earlier attribution of the campaign to the Llama 3 family, while also turning up evidence that Mistral-family models were involved.
What makes this notable isn't that a propaganda operation used AI - that's assumed at this point. It's that the operators left behind their actual editorial playbook, in plain text, on their own websites, letting researchers reconstruct the assembly line rather than just guess at it. That kind of leak also gives attribution efforts something more concrete than stylistic hunches to work with.
This is one paper examining one campaign's leftovers, not a general rule about how sloppy AI propaganda operations are - but it's a useful reminder that automating disinformation at scale still means leaving a paper trail somewhere.