Nearly half of all participants in a new controlled experiment failed to correctly identify more AI-generated comments than they falsely flagged as bots - even though the task was explicitly framed as a bot-detection challenge. The findings, produced by cybersecurity company Surfshark in collaboration with master's students at Malmö University, point to a vulnerability that is less about technical knowledge and more about the way human emotion short-circuits critical judgment. The experiment arrives at a moment when automated accounts have become a structural feature of online political life, not a fringe nuisance.
What the Experiment Actually Tested
The "Bot or Not" simulation was built as an interactive exhibit for the UNFOLD exhibition at Milan Design Week. It places participants inside a mock social media comment section and gives them 120 seconds to identify 10 bot-written comments spread across four discussion topics. The design is deliberately simple - no technical tools, no metadata, just the text of the comments themselves, which is precisely how the vast majority of ordinary users encounter content online.
Of the 710 people who participated, only 53% correctly identified more bots than they misidentified real humans. That leaves 47% who, by the study's own measure, failed the task. These were not passive or inattentive users. They had opted into an exercise explicitly described as a bot-detection challenge, which means the actual gap in everyday, unconsidered scrolling is likely wider still.
The four topics in the simulation were not chosen arbitrarily. Two were emotionally neutral: the environmental footprint of data centres and the long-running internet argument over pineapple as a pizza topping. Two were emotionally charged: immigration policy and women's rights. The contrast was the experiment's sharpest finding. On data centres, participants detected 71% of bots at a 76% accuracy rate. Pineapple on pizza produced similar results - 64% detection, 69% accuracy. The moment the subject matter carried political or emotional weight, both numbers collapsed. On immigration, detection fell to 54% with 63% accuracy. On women's rights, detection dropped to just 49%, with accuracy at 61%. Participants were simultaneously missing more bots and wrongly accusing more genuine human commenters of being machines.
Emotion as a Structural Weakness
The mechanism behind this performance gap is not mysterious, even if its implications are underappreciated. When a topic triggers strong feelings - anger, fear, moral outrage - cognitive resources that would ordinarily be available for evaluating the plausibility of a source get redirected toward processing the emotional content itself. A comment that aligns with a reader's existing convictions feels authentic because it resonates, not because it displays markers of genuine human authorship. A comment that provokes outrage reads as a real person being wrong, not as a machine executing a persuasion strategy.
This is precisely the vulnerability that sophisticated bot operations are designed to exploit. The goal of politically motivated automated accounts is rarely to spread information that is obviously false. It is to amplify emotional signals - to make a debate feel more polarised, a consensus more fragile, or a fringe position more mainstream than it actually is. Surfshark's Research Lead Luís Costa has described the core takeaway in direct terms: the biggest blind spot the experiment exposed was emotion. What users need, he argues, is not sharper reading skills in the traditional sense, but a clearer awareness of when their own reactions are making them easier to mislead.
The scale of the problem provides context for why individual detection failures matter. Surfshark's earlier research found that major platforms collectively remove more than 6.3 billion fake accounts each year - roughly 47 times the number of children born globally in the same period. That figure represents accounts caught and removed. It says nothing about the accounts that remain active, the comments that were already read, or the impressions that were already formed before any moderation occurred. Separate industry estimates suggest that bot-driven amplification accounts for around 23% of political discourse on X during election periods, a proportion large enough to meaningfully distort the apparent shape of public opinion.
The Age Factor and What It Reveals
The study also identified a pronounced drop in detection ability at around the age of 40. Participants under 20 were the strongest performers, finding nearly 65% of bots with an accuracy rate above 71%. Results held broadly steady through the 20s and 30s, then fell sharply for the 41-to-50 age bracket, where detection dropped to 42% and accuracy fell to 59%. Users over 50 performed only marginally better than that cohort.
The generational pattern is worth reading carefully. Younger users grew up forming their social instincts in environments already saturated with algorithmic content, automated accounts, and the ambient awareness that not everything online originates from a human. That familiarity appears to confer a practical advantage when it comes to recognising the subtle stylistic flatness of machine-generated text. Older users, whose baseline expectations of online communication were set in an earlier era, may be applying a framework of assumed human authorship that the current information environment no longer warrants.
This does not mean younger users are immune. Their advantage existed under controlled conditions, on a task they knew was about bots. The real question is whether that alertness persists during ordinary, emotionally engaged browsing - and the emotional-topic data suggests it may not, regardless of age.
What This Means for Public Discourse
The "Bot or Not" experiment is a small study with a specific methodology, and its findings should be read as indicative rather than definitive. But the direction of the results is consistent with a broader pattern that researchers studying computational propaganda have documented across multiple platforms and election cycles. Automated accounts work not by fooling users into believing a single fabricated claim, but by gradually reshaping the perceived landscape of opinion - making certain positions appear more numerous, more energised, or more socially acceptable than they are.
The implication for anyone who consumes political content online is uncomfortable but straightforward. The moments when a comment section feels most alive, most contested, and most emotionally urgent are precisely the moments when automated amplification is most likely to be operating - and when the human capacity to detect it is at its lowest. Awareness of that dynamic does not require any specialised skill. It requires only the habit of pausing to ask whether a reaction that feels very immediate and very real was, in fact, engineered to feel that way.
The "Bot or Not" simulation is publicly available at botornot.one. It takes two minutes to complete and produces a personal score benchmarked against the original 710 participants. The number on the screen matters less than what the experience of playing it tends to reveal: that confidence and accuracy are not the same thing, and that the topics people feel most certain about are often the ones where their judgment is most exposed.
