Nearly half of participants in a controlled experiment failed to identify more AI-generated social media comments than they wrongly flagged as bots - even though they considered themselves digitally literate. The finding, from a study conducted by cybersecurity company Surfshark in collaboration with master's students at Malmö University, points to a vulnerability that has less to do with reading ability and more to do with how human emotion short-circuits critical judgment. The experiment's implications extend well beyond a browser-based exercise: bots now account for an estimated 23% of political discourse on X during election periods, according to recent industry figures.
What the Experiment Actually Measured
The "Bot or Not" simulation was built as an interactive installation for the UNFOLD exhibition at Milan Design Week. Interaction Design master's students at Malmö University designed it as a timed challenge: participants enter a simulated social media comment section and have 120 seconds to identify 10 bot-written comments spread across four discussion topics. The setup deliberately mixed emotionally neutral subjects with politically charged ones, and that contrast produced the study's most striking results.
On data centres - a dry, technical topic - participants identified 71% of bots with a 76% accuracy rate. The pineapple-on-pizza debate, trivial but mildly entertaining, yielded 64% detection and 69% accuracy. Both are reasonable results. Then the topics shifted. On immigration, detection dropped to 54% and accuracy fell to 63%. On women's rights, detection collapsed to just 49%, with accuracy at 61%. Participants were not only missing more bots on those topics - they were also misidentifying more real humans as machines. The emotional charge of the subject did not sharpen their attention. It degraded it.
Of the 710 total participants, only 53% correctly identified more bots than they misidentified humans. The remaining 47% failed the core task entirely.
Emotion as a Mechanism of Manipulation
The pattern revealed here is not incidental. It reflects a deliberate design logic in how modern bots operate. Automated accounts amplifying political content are not primarily engineered to sound neutral or academic. They are built to provoke, to mirror existing grievances, to confirm what an emotionally activated reader already suspects. When a comment says something that feels familiar, urgent, or outrageous, the instinctive reaction is engagement - not scrutiny.
Surfshark's Research Lead Luís Costa frames the core problem this way: the experiment exposed emotion as the primary blind spot. When a discussion becomes heated, the mental process people rely on to flag suspicious behaviour is effectively overridden. The implication is that traditional media literacy - knowing how to check sources, recognising sensationalist language, understanding how algorithms work - is insufficient on its own. What users need is an awareness of their own psychological state at the moment they encounter contested content. A reader who knows they are already irritated or invested in a topic is a reader who is more vulnerable, not less.
This is not a problem that technology alone can solve. Platforms remove fake accounts at enormous scale - Surfshark's earlier research found that major platforms collectively remove more than 6.3 billion fake accounts per year, roughly 47 times the annual number of births worldwide. Yet the accounts keep appearing, the technology generating them keeps improving, and the emotional levers they are designed to pull remain just as effective.
A Generational Pattern Worth Taking Seriously
The study also identified a clear performance drop at around age 40. Participants up to age 20 were the most reliable at identifying bots, detecting nearly 65% of them with an accuracy above 71%. Results held reasonably steady through the 20s and 30s, then dropped sharply for the 41 to 50 age group, where detection fell to 42% and accuracy to 59%. Users over 50 performed only marginally better than that bracket.
This pattern inverts a common assumption. Older users are often assumed to bring more life experience and scepticism to online spaces. What the data suggests instead is that younger users - who grew up navigating algorithmically driven social media and have developed an intuitive, fast-processing familiarity with its conventions - may be better calibrated to notice when something feels off. The more revealing question is whether that edge will persist as bot-generation technology continues to advance, or whether it simply reflects a familiarity gap that will close as older cohorts spend more time in those environments.
What This Means for the Information Environment
The broader context makes the experiment's findings harder to dismiss as a novelty. Automated amplification of political messaging during elections is not a theoretical risk - it is a documented, recurring feature of how information now moves on major platforms. When nearly half of self-identified savvy users cannot reliably distinguish machine-written comments from human ones, the question of whether public discourse online reflects actual public opinion becomes genuinely difficult to answer.
The "Bot or Not" simulation is now publicly accessible at botornot.one. It takes a few minutes to complete and provides an individual result against the baseline of the original 710 participants. That comparison is useful, but the more durable value of the exercise is what it demonstrates about the conditions under which critical thinking tends to fail - not through ignorance, but through emotional activation. Recognising that mechanism is the first step toward being less susceptible to it.
