A new global study from Surfshark indicates that social media users struggle to identify AI-driven bots, with friendly and agreeable automated accounts achieving significantly higher deception rates than aggressive or confrontational ones.
The research evaluated 1,722 participants worldwide to measure their ability to distinguish human comments from AI-generated outputs in simulated social media environments. Overall, participants successfully detected only 40% of the deployed bots, exposing widespread vulnerability to automated social engineering.
According to the findings, negative and confrontational bots were flagged far more frequently because their hostile tone immediately triggered user suspicion. Participants successfully identified 50.2% of negative AI-generated bots during the testing period.
In contrast, when automated accounts adopted a positive, agreeable, and logical persona, detection rates dropped sharply to 38%. Luís Costa, Research and Insights Lead at Surfshark, noted that online users naturally focus on hostile accounts while letting friendly actors bypass scrutiny. “On social media, everyone notices the angry trolls. That’s why the angry trolls are the ones who often get caught,” Costa stated.
Costa further explained that sophisticated manipulation campaigns leverage this behavioral tendency by deploying agreeable personas that validate real users’ opinions and artificially inflate discussion volumes. “Often, AI-powered accounts used in sophisticated manipulation campaigns are agreeable, logical, or simply unremarkable. They can support real users’ opinions and just inflate the number of comments in the discussion to make a minority view look like everyone feels the same way. Seldom do people report such accounts,” Costa added.
The study also highlighted demographic and contextual variables affecting bot detection. Participants demonstrated higher accuracy when evaluating lighthearted topics compared to serious societal issues like immigration, where positive bots proved exceptionally difficult to spot. Additionally, detection capability declined steadily with age, showing the lowest scores among participants over the age of 50.