How a Russian Fake News Site Used AI to Mass-Produce Persuasive Lies

A covert Russian propaganda site used generative AI to mass-produce misleading articles, without sacrificing their persuasive power. A new study shows how easily these tools can scale up and sharpen the spread of false narratives.

Research: Generative propaganda: Evidence of AI’s impact from a state-backed disinformation campaign. Image Credit: Roman Samborskyi / Shutterstock

A study of the use of AI by a Russian-backed propaganda outlet shows how AI allows propagandists to increase their production capacity without any loss in persuasive power.

In December 2023, following a BBC report identifying DCWeekly.org as part of a Russian disinformation operation, researchers at Clemson University's Media Forensics Hub published an investigation revealing that the site DCWeekly.org was a Russian propaganda outlet, part of a broader network disseminating pro-Kremlin and anti-Ukrainian narratives.

Morgan Wack and colleagues found that prior to September 20, 2023, much of the content on the site was simply lifted from other right-leaning outlets, including Russian state media. After that date, however, the stories were generally AI-scored, selected, and rewritten using OpenAI’s GPT-3 model, allowing the site to use a broader range of sources —including translated Russian state media and mainstream outlets like Fox News— but with the tone and emphasis tweaked to better suit likely aims of the propagandists.

The authors examined 22,889 articles published on the site before and after the shift, which followed a period of declining output, but before the site's true motives were exposed. By leveraging the timing of its adoption, the research team shows how AI allowed the propagandists behind the site to more than double their rate of publication, as compared to their most active pre-AI period.

The use of AI by DCWeekly also corresponded with an increase in the breadth of topics covered on the site, which ranged from Russia's purported successes in Ukraine to the tyranny of gun control legislation in the United States, as measured through topic modeling and natural language inference classification.

As a final step, the authors conducted a survey of 880 American adults recruited through the online survey platform Prolific and found that the content in the post-adoption period maintained the same level of persuasiveness and credibility as the pre-AI period.

The study suggests AI may help propagandists flood online discourse with divisive content, even if individual articles do not become more persuasive. After DCWeekly’s exposure, similar AI-assisted fake news sites emerged, suggesting a scalable disinformation template. According to the authors, immediate action should be taken to mitigate the influence of AI-assisted propaganda campaigns.