Nearly half of people tested in a new simulation cannot reliably tell AI-generated comments from those written by real humans - and the problem gets measurably worse the moment the conversation turns political. A study conducted by cybersecurity company Surfshark, built in collaboration with Interaction Design master's students at Malmö University, tested 710 participants on their ability to identify bot-written content across four distinct discussion topics. The results suggest that emotional context, not technical literacy, is the primary variable determining whether someone flags a bot or misses one entirely.
What the Experiment Actually Tested
The "Bot or Not" simulation places participants in the role of a content moderator inside a fictional social media comment section. Players have 120 seconds to identify 10 bot-written comments spread across four topics. Two were designed to be emotionally neutral - data centres and the question of whether pineapple belongs on pizza. The other two were deliberately charged: immigration and women's rights.
The gap between those two categories is where the study's central finding lives. On data centres, participants correctly identified 71% of bots with a 76% accuracy rate. The pineapple debate produced similar results - 64% detection, 69% accuracy. Both figures suggest that, in the absence of emotional pressure, people retain a reasonable instinct for what sounds artificially generated.
That instinct deteriorated sharply once the topics became politically sensitive. On immigration, detection dropped to 54% and accuracy to 63%. On women's rights, detection fell to just 49% - effectively a coin toss - with accuracy at 61%. Participants were not only missing more bots; they were also wrongly flagging more real humans as machines. The simulation was exposing something beyond reading comprehension: it was surfacing how emotional engagement corrupts the cognitive process of evaluation itself.
The Generational Pattern and What It Reveals
The study identified a pronounced drop in detection ability at around the age of 40. Participants under 20 were the strongest performers across the dataset, correctly identifying nearly 65% of bots with an accuracy rate above 71%. Detection held at comparable levels through the 20s and 30s, then fell sharply for the 41-50 bracket, where detection slipped to 42% and accuracy to 59%. Users over 50 performed only marginally better than that cohort.
This does not straightforwardly indicate that older users are less digitally capable. More plausible explanations include differences in platform familiarity, variation in how different age groups engage with politically contentious content online, and the possibility that younger users - having grown up alongside algorithmically curated feeds - have developed a more reflexive skepticism toward the comment sections they scroll through daily. The pattern warrants further research, but the data point is stark enough to take seriously.
According to Surfshark's Research Lead Luís Costa, the core vulnerability exposed by the experiment is not a deficit in reading skills or conventional media literacy. The real problem is emotional: when a debate becomes heated, it effectively disables the mental process people rely on to assess whether a given comment is authentic. What the findings argue for, Costa suggests, is a greater awareness of personal emotional vulnerability rather than more sophisticated textual analysis.
The Scale of the Problem Beyond the Simulation
The experiment is a contained exercise, but the environment it simulates is not. Surfshark's own earlier research found that major platforms collectively remove more than 6.3 billion fake accounts each year - a figure that amounts to roughly 47 times the number of children born globally in the same period. Industry estimates place bot-driven amplification at around 23% of political discourse on X during election cycles. These are not fringe phenomena. They represent a structural feature of how political and social conversation now functions online.
The underlying technology is also not static. The same advances in large language models that have made AI-generated text more fluent and contextually coherent in professional settings have made bot-written social media content harder to distinguish from genuine human expression. A comment that would have read as clunky or formulaic a few years ago may now be indistinguishable from what a real person might write in a moment of genuine frustration or conviction. That convergence is precisely what makes emotional manipulation such an efficient vector: the bots do not need to be perfect, they only need to be plausible enough when the reader is already agitated.
What Comes Next for Platforms and Users
Platform-level moderation removes billions of fake accounts annually, but the gap between creation and removal represents a meaningful window of influence - particularly during elections, public health crises, or any period when coordinated narrative shaping carries real-world consequences. Detection tools powered by machine learning exist and continue to improve, but they face an adversarial dynamic: the same techniques used to identify bot-generated content can, in principle, be used to train the next generation of bots to evade detection.
For individual users, the practical implication of the Surfshark study is uncomfortable but clarifying. The content most likely to manipulate is the content most likely to feel urgent, morally important, or emotionally resonant. That is not an accident of design - it is the design. Recognising that the feeling of certainty triggered by a provocative comment is itself a potential point of vulnerability is a more durable form of protection than any checklist of stylistic red flags.
The "Bot or Not" simulation is publicly available at botornot.one. Running through it takes roughly two minutes. Whether the score it returns prompts a genuine reassessment of one's own online instincts is, of course, another question entirely.
