Nearly half of participants in a new cybersecurity experiment could not correctly identify AI-generated comments on social media more often than they falsely flagged real human posts - and the determining factor was not age or digital literacy, but emotional temperature. The study, conducted by cybersecurity company Surfshark in collaboration with master's students from Malmö University, exposed a vulnerability that no amount of technical know-how appears to reliably fix: when a conversation gets charged, human judgment degrades fast.
What the Experiment Actually Tested
The "Bot or Not" simulation was built as an interactive installation for the UNFOLD exhibition during Milan Design Week. Players enter a mock social media comment section and have 120 seconds to identify 10 bot-written comments spread across four discussion topics. The format is deliberately low-tech in its ask - read, judge, click - which is precisely what makes the results uncomfortable.
Of the 710 participants tested, only 53% correctly identified more bots than they wrongly accused real humans of being. The remaining 47% effectively failed. The study did not select for naïve users. Many participants considered themselves digitally literate, which makes the failure rate harder to dismiss as a problem belonging to someone else.
The four topics were split by emotional charge. Data centres and the pineapple-on-pizza debate served as low-stakes, "cold" discussions. Immigration and women's rights were the "hot" topics - politically loaded, prone to strong reactions. On data centres, participants detected 71% of bots with a 76% accuracy rate. On women's rights, detection collapsed to 49% and accuracy to 61%. The bots were not better written on the emotional topics. The readers were simply worse at reading them.
Why Emotion Disarms the Very Instinct That Should Protect Us
According to Surfshark's Research Lead Luís Costa, the core finding is not about reading comprehension or familiarity with AI writing patterns. It is about the way strong emotion commandeers the cognitive resources people normally use to assess credibility. When a comment triggers agreement, outrage, or recognition, the mental step of asking "is this even a real person?" tends to get skipped entirely.
This is not a new psychological mechanism - it is the same principle that makes emotionally resonant misinformation more viral than dry corrections. What the experiment adds is a direct, measurable demonstration of how that mechanism plays out in the specific context of bot detection. The problem is not that people cannot recognise AI-generated text in principle. It is that they stop trying to when the content already feels meaningful to them.
The implication for disinformation campaigns is direct. Bots seeded into politically charged discussions are not just present where the audience is largest - they are present where the audience's critical defences are lowest. Industry estimates suggest bot-driven amplification accounts for around 23% of political discourse on X during election seasons, a figure that lands differently once you understand how compromised human detection becomes in exactly those contexts.
The Generational Gap and What It Does - and Does Not - Explain
The study identified a pronounced drop in bot-detection ability at around age 40. Participants under 20 performed best, identifying nearly 65% of bots with an accuracy rate above 71%. Performance held reasonably steady through the 20s and 30s, then fell sharply for the 41 to 50 age bracket, where detection dropped to 42% and accuracy to 59%. Users over 50 showed only marginal improvement over that cohort.
The tempting interpretation - that younger users are simply more digitally fluent - does not fully hold. Surfshark's own earlier data shows that major platforms collectively remove more than 6.3 billion fake accounts every year, a volume roughly 47 times the number of babies born worldwide annually. The bots being created today are not the clumsy, obviously mechanical outputs of a decade ago. They are trained on vast pools of human writing, calibrated to mirror the cadence and vocabulary of authentic conversation. Younger users may be better at pattern recognition, but they are facing a moving target that is converging rapidly with the human writing it mimics.
What the generational pattern more likely reflects is a combination of factors: greater exposure to diverse online communication styles among younger cohorts, faster adaptation to AI-era writing norms, and possibly a higher baseline suspicion of content encountered online. None of those advantages are fixed, and none of them held up reliably under emotional pressure in any age group.
The Broader Stakes of Getting This Wrong
The scale of synthetic social media activity is not a background nuisance - it is an infrastructure problem. Billions of fake accounts are manufactured, deployed, and removed annually, only to be replaced by new ones. The technology behind them improves continuously. The social media platforms tasked with removing them are operating at a deficit that their own deletion figures make plain.
Individual users are not equipped to serve as a meaningful last line of defence, and the "Bot or Not" experiment makes a useful case for why designing systems around that expectation is a structural error. Costa's conclusion is pointed: the solution is not sharper textual analysis but better self-awareness about emotional vulnerability. Knowing that your critical faculties are likeliest to fail you precisely when a conversation matters most to you is not a comfortable insight, but it may be the most actionable one the study produces.
The simulation is publicly accessible at botornot.one, and the few minutes it takes to complete are a more honest self-assessment than most users are likely to seek out voluntarily. The score matters less than the conditions under which it drops.
