Surfshark's agreeable bot personas evaded detection 62% of the time

Converted into miss rates: positive, agreeable personas went unflagged 62% of the time against 49.8% for the hostile ones, and the miss rate across the whole test was 60% ^[1][2][3].

User reporting inherits those numbers, and so does the awareness training that tells staff to watch for accounts behaving strangely. Training describes the persona participants caught half the time ^[4].

On the attacker's side the adjustment is free. Surfshark said an account that stops overusing emojis becomes twice as hard to identify ^[7]. On its own published figures the detection rate falls by a factor of about 1.7 and the unflagged share rises by about 1.6 ^[4].

"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," said Luis Costa, Research and Insights Lead at Surfshark ^[15].

The error ran both ways on the harder subject. On women's rights participants missed more bots than on a light topic such as pineapple on pizza, and they also wrongly accused more real humans ^[8]. Immigration reordered the result: negative bots were still the easiest to catch there, but positive bots were harder to spot than in any of the other three topics ^[9]. Neutral bots sat near the bottom for detection across most topics ^[10].

Threads users had the highest detection rate, on what Surfshark notes is a small sample, with X users close behind and TikTok and Facebook users well behind both ^[12]. Participants who use social media almost all the time caught close to half the bots they saw, while people who do not use it at all caught roughly a third ^[13].

The test measured whether participants could separate human comments from AI-generated ones in a simulated social media setting ^[1]. Surfshark's researchers argue the agreeable, logical bot is a bigger privacy and security concern than the confrontational one precisely because it does not set off suspicion ^[14]. The step from a missed label to a private chat, a handed-over detail or a clicked link is their argument from the detection gap, and the study is Surfshark's own ^[1].