Propaganda Forensics: Recovering the Generation Pipeline of an AI-Driven Influence Campaign

In the modern information landscape, a new kind of noise has emerged, born from the rapid advancement of artificial intelligence. Large language models, the powerful computer systems capable of generating human-like text, have lowered the barrier for creating news articles at an industrial scale. While these tools are often celebrated for their utility, they possess a dual nature: they can be used to rewrite existing stories into a form designed to manipulate public opinion. This phenomenon, sometimes called "slopaganda," involves taking real news and reshaping it to fit a specific, often misleading, narrative. The challenge for researchers is no longer just spotting obvious lies, but distinguishing between genuine journalism and content that has been algorithmically altered to look authentic. To understand how this manipulation works, scientists must act as forensic analysts, examining the digital fingerprints left behind by the machines that create the text.

A team of researchers from institutions including Sorbonne University and Airbus recently turned their attention to a specific, large-scale influence campaign known as Storm-1516. This operation, which targeted French audiences, operated a network of eighty-four websites that impersonated legitimate news outlets like Le Monde and France Télévisions. These sites published thousands of articles that appeared to be standard news but were actually part of a coordinated effort to spread pro-Russian narratives. The researchers gathered 2,646 of these articles into a new collection they named PROPAGIA. To understand what made these texts different from real journalism, they compared them against a control group of 2,385 articles written by humans for a major French daily newspaper, Ouest-France, covering the same time period. The goal was to determine not only if the texts were written by machines, but to uncover the specific instructions the machines were given and to identify the family of artificial intelligence models responsible.

The first step in the investigation was to analyze the style and tone of the writing. The researchers found that the articles in the PROPAGIA collection were significantly more vague and subjective than the human-written articles. Where a real journalist would typically cite specific sources, name experts, or provide concrete details, the AI-generated texts relied heavily on general statements and emotional language. The propaganda articles were also far more negative, often building a sense of fear or exaggeration around events. They cited external voices much less frequently, substituting the author's own assertions for verifiable facts. This pattern suggested a deliberate strategy to replace factual reporting with opinion, a technique that makes the content harder to fact-check and easier to use for emotional manipulation.

The most direct evidence of machine involvement came from a surprising source: the text itself. In fifty of the eighty-four websites, the researchers found that the artificial intelligence had accidentally left behind parts of its own instructions. These "leaks" appeared as notes or checklists embedded within the published articles, revealing the exact commands the operators had given the machine. One such document listed a ten-point editorial specification, instructing the model to rewrite stories while avoiding mentions of other media, condemning specific political leaders, and framing the Russian president in a positive light. These instructions explained the patterns seen in the text: the lack of sources, the negative tone, and the specific political slant were not random errors but the result of a strict, automated recipe. The presence of these leaks confirmed that the content was not merely written by a human with a bias, but was generated by a machine following a rigid set of rules.

To determine which specific type of artificial intelligence was used, the researchers employed a method that tested how different models reacted to the text. They asked several different AI systems to rewrite the articles, observing how much each system changed the original wording. The logic was that if an AI model had originally written a text, it would be less likely to change its own work significantly when asked to rewrite it, compared to how it would change a text written by a human. The results showed a clear pattern: the models suspected by intelligence agencies to be behind the campaign, specifically those from the Llama 3 family, altered the propaganda articles the least. Interestingly, models from the Mistral family showed a similar pattern, suggesting that the campaign might have used a mix of related technologies or that these different families share underlying characteristics.

The study concluded that the Storm-1516 campaign was a sophisticated operation that relied on a pipeline of generative AI to mass-produce propaganda. By analyzing the text for signs of vagueness, finding leaked instructions, and testing how different models rewrote the content, the researchers were able to reconstruct the generation process. They found that the campaign used a specific set of instructions to force the AI to produce negative, source-less narratives that mimicked the style of real news. While the researchers could not pinpoint the exact single model used, their evidence strongly suggested that the Llama 3 family of models was central to the operation, potentially alongside Mistral models. This forensic work provides a blueprint for how to detect and understand future AI-driven influence campaigns, moving beyond simple detection to understanding the mechanics of how the manipulation is engineered.