Five Myths About Influence Operations: What 25 Million Tweets Across Seven State Campaigns Reveal

This study challenges five widely held "stylized facts" about state-backed influence operations (IOs). By analyzing a comprehensive archive of over 25 million tweets from seven state-attributed campaigns—including those from Russia, Iran, Venezuela, and Bangladesh—the author re-evaluates how these operations function. Using a rigorous, pre-registered protocol that incorporates matched organic baselines, false-discovery control, and placebo tests, the research systematically dismantles the prevailing narrative that these operations are sophisticated, adaptive, and self-amplifying entities.

The paper tests five common beliefs: that IOs are monolithic armies, that they succeed by weaponizing moral-emotional language, that they manufacture their own virality through internal amplification, that they are sophisticated learners that optimize against feedback, and that they are indistinguishable from ordinary users. The results show that these operations are actually narratively segregated, thinly staffed desks. They do not rely on moral-emotional contagion to drive engagement, and internal amplification accounts for a negligible fraction of their total reach. Furthermore, these operations are largely scripted rather than adaptive; when they appear to respond to feedback, they typically mean-revert toward a baseline rather than optimizing for success.

While the linguistic "troll" fingerprints of the 2016 era have largely aged out, the study finds that these operations remain detectable through their coordination patterns. Even as their content has become more indistinguishable from ordinary users, the accounts within these campaigns coordinate 7–70 times more tightly than matched organic users. This suggests that the primary signature of an influence operation has migrated from the content itself to the underlying coordination structure. The author warns that while these specific operations are not currently optimizing via feedback, the future integration of large language models could enable adversaries to overcome these limitations, making the current findings a baseline rather than a permanent state of affairs.