The Silent Farm: Bots, Public Perception, and the Invisible Beneficiary
There is a moment when a number stops being a statistic and becomes a symptom. That moment, as far as the internet is concerned, has already passed. It passed without newspaper headlines, without political declarations, without any institutional response remotely proportional to its gravity.
In 2024, for the first time in a decade, automated traffic surpassed human activity online. More than half of everything circulating on the internet — 51 per cent, according to Imperva's annual Bad Bot Report — was produced by no human hand, observed by no human eye, driven by no human intention. It was produced by machines that imitate people. Machines that comment, distribute likes, create accounts, spread messages, simulate consensus, and construct — with the tireless patience of a mechanism that never sleeps — a parallel reality in which no one knows with any certainty who they are actually talking to.
This is not a metaphor. It is a cybersecurity report. And its implications extend far beyond the world of IT.
Before 51%: How We Got Here
The internet was not born with bots. It was born with people who dreamed of a global network of knowledge — a space where ideas might travel freely and conversation transcend borders. The first bots arrived quietly, in the shadow of that dream: simple programmes that indexed pages, dispatched automated notifications, responded to elementary commands.
They were tools. Neutral, in principle.
The transformation was gradual. Social media platforms came to understand that engagement — not truth, not the quality of conversation, but raw emotional reaction, any form of activity at all — was their currency of exchange. And once someone grasped that bots could generate engagement on an industrial scale, they became instruments of power.
By 2023, nearly half of all global internet traffic — 49.6 per cent, according to the same Imperva research — was already coming from automated sources. The highest level recorded since the company began monitoring the phenomenon a decade earlier. A year later, the 50 per cent threshold was crossed. This trend is neither accidental nor reversible through individual vigilance alone.
There is a further acceleration, more recent and more troubling: generative artificial intelligence. ChatGPT, Gemini, open-source models accessible to anyone with a laptop — all of these have dramatically reduced the cost and complexity of building a convincing bot. Where once you needed specialist programmers and considerable resources to construct a credible network of false accounts, today a teenager with an LLM and an afternoon to spare can deploy a small army of digital entities capable of producing coherent, nuanced text that is genuinely difficult to distinguish from the human kind.
"Generative AI alone is not more dangerous than bots," said Kathleen Carley, a researcher at Carnegie Mellon University. "It's bots plus generative AI." It is a formulation worth repeating, because it captures precisely the nature of the mutation: this is not a new threat. It is an old threat amplified exponentially by new tools.
The Anatomy of Invisible Traffic
Not all bots are created equal, and not all of them are malevolent. Cybersecurity literature draws a fundamental distinction: good bots versus bad bots.
Good bots are the crawlers used by search engines — Googlebot, Bingbot — along with news aggregators, AI assistants that index content in order to cite or summarise it, and server monitoring tools. Without them, the internet as we know it — its entire architecture of search and discovery — would simply not function. These bots are a legitimate and necessary part of online traffic.
Bad bots are something else entirely. They come in three tiers, classified by their degree of sophistication: simple, moderate, and advanced. The simple ones execute elementary tasks — sending spam, filling in forms, generating fraudulent advertisement clicks. Moderate bots are capable of mimicking human behaviour convincingly enough to pass basic filters. Advanced bots — the fastest-growing category, propelled precisely by AI — simulate real browsers, vary their behaviour, analyse previous failures, and adapt dynamically to avoid detection.
In 2024, malicious bot activity rose for the sixth consecutive year, reaching 37 per cent of all internet traffic, up from just over 30 per cent in 2023. That means more than a third of everything that circulates online is, in essence, an attack in progress — whether we are talking about aggressive data scraping, illegal account takeovers, advertising fraud, or, most relevant to the matter at hand, the manipulation of public opinion.
There is also an economic ecosystem sustaining all of this infrastructure. It is called Bots-as-a-Service — BaaS — and it operates exactly like any other industry: suppliers, clients, pricing plans, technical support. Someone wants to flood a competitor's comments with negative reviews? There is a package for that. Someone wants to manufacture the appearance of a political majority? There is a package for that, too. The market is mature, competitive, and, for the most part, comprehensively illegal — yet extraordinarily difficult to prosecute, and almost never actually pursued.
Social Media as Theatre of Operations
If web traffic in general is already dominated by bots, social media is the terrain where the consequences of that domination become genuinely dangerous. Not because the absolute volume of automated traffic is highest there, but because that is where public opinion is formed, where collective perceptions crystallise, where political reputations are built — and destroyed.
Academic studies estimate that between 9 and 15 per cent of all active accounts on social media platforms are automated. At first glance, this sounds like a negligible minority. But that view overlooks a fundamental truth about how bots actually operate: they do not sleep. They do not tire. They have nothing else to do and no days off. A single bot can produce hundreds — or thousands — of comments a day; the most prolific human user might manage a few dozen. In coordinated campaigns, the volume of bot-generated comments can exceed that of real users by several orders of magnitude.
The consequence is that what appears to be a human conversation may in reality be saturated with automated messages. And the person reading those comments — who sees dozens or hundreds of apparently independent voices endorsing the same idea, attacking the same individual, amplifying the same fear — has no simple instrument with which to distinguish genuine consensus from manufactured consensus.
This mechanism has been documented in detail across multiple major electoral cycles. Emilio Ferrara at USC studied bot operations during the 2016 and 2018 American presidential and midterm elections, as well as the 2017 French presidential election, demonstrating that these accounts were used systematically to manipulate public discourse and amplify disinformation. In 2024, troll farms continued to operate at scale, targeting both American political parties and inflaming divisions around race, gender, and immigration — not necessarily in order to deliver victory to any particular candidate, but to deepen polarisation, erode trust in institutions, and render genuine public conversation impossible.
There is also a European case that is frequently overlooked in discussions of digital manipulation. An audit conducted on TikTok during the 2024 German regional elections found that new users — those with no declared political preferences — were three to four times more likely to encounter content from a far-right populist party than from mainstream parties, within just three weeks of using the platform. This was not necessarily the direct effect of bots, but of algorithms that amplify divisive content. Bots and algorithms frequently work in the same direction, even when their mechanisms differ.
Perception as Prey
There is a concept in social psychology known as social proof: the tendency of human beings to regard a behaviour or opinion as correct when a large number of others appear to share it. It is an evolutionarily rational mechanism in a world where information is scarce and costly. When everyone in the village is running, there is probably a reason to run.
On social media, this mechanism is exploited systematically. A post with ten thousand likes appears more credible than one with fifty — even if those ten thousand likes were purchased or generated automatically. A comment with dozens of agreeing replies seems to reflect genuine consensus — even if half of those replies were produced by coordinated bots. From the perspective of the ordinary user, reality and its simulacrum are indistinguishable.
The cumulative effect is what researchers call the manipulation of collective perception. You do not persuade people to believe something new; you persuade them that what they already believe is a marginal minority view, that their opinion is unsupported, that everyone else thinks differently. The result is the suppression of authentic voices and the artificial amplification of false ones.
In 2025, the Centre for Emerging Technology and Security at the Alan Turing Institute published a report on AI influence operations in electoral contexts. The conclusion is nuanced, but far from reassuring. There is no conclusive evidence that AI-generated disinformation directly altered the outcome of the 2024 American presidential election — but deceptive AI-generated content did influence public discourse, amplify harmful narratives, and deepen political polarisation. The distinction between "changed the result" and "distorted the conversation" is less consoling than it sounds. A democracy in which public conversation is systematically distorted does not function as a democracy, regardless of whether the final vote happens to reflect the majority's preference.
The Detection Problem: When Machines Imitate People Better Than People Do
There is a comfortable temptation to believe that sophisticated bots are easily detected — that clear tell-tale signs betray them, that an attentive user can spot them with practice. This belief is increasingly difficult to sustain.
Current detection methods for inauthentic coordinated behaviour identify clusters of accounts that post in synchronised patterns, amplify the same users, or distribute identical sets of links, images, and hashtags — patterns that give away coordination. These methods are valuable, but they have serious limitations. Advanced bots, trained on modern language models, vary their behaviour deliberately, introduce noise to break detectable patterns, and produce content diverse enough to avoid triggering automated alerts.
More troubling still: researchers at Indiana University's Observatory on Social Media have found that current AI content detectors cannot reliably distinguish LLM-powered social bots from genuine human accounts. The machine writing the bot's comments has advanced faster than the machine tasked with detecting them. It is an asymmetric arms race — one in which the attacker holds the inevitable advantage of choosing the terrain.
This means the problem can no longer be resolved by technical means alone. No algorithmic filtering, however sophisticated, will eliminate automated activity from social platforms as long as the cost of producing a convincing bot continues to fall and its quality continues to rise.
The Institutional Void: Or, Why Nothing Has Been Done
We are confronted with a deeply troubling paradox. We have detailed reports. We have robust academic research. We have concrete evidence that bot infrastructure influences public discourse and, by extension, democratic processes. And yet, at state level — in Romania, certainly, but not only there — not a single concrete, systematic measure has been taken.
There is no national authority with an explicit mandate to monitor automated traffic on social platforms. There is no legislation requiring platforms to limit or declare coordinated automated activity. There is no mechanism for verifying real identity for social media accounts participating in public conversations with political stakes.
Three hypotheses present themselves for this inaction, and they are not mutually exclusive.
The hypothesis of technical incompetence. Most policymakers do not genuinely understand the phenomenon. They hear the word "bots" and picture something from a science fiction film — not an industrial infrastructure operating on servers distributed across the global cloud. This incompetence is not morally culpable, but it is practically dangerous.
The hypothesis of regulatory capture. Social media platforms are private companies with formidable lobbying resources. Any meaningful regulation of bots would require changes to their business model — less engagement, less traffic, less advertising revenue. The industry's resistance to regulation has been both consistent and effective.
The hypothesis of direct benefit. This is the most uncomfortable hypothesis, but perhaps the most honest. Bot farms are useful to political actors who wish to manufacture the appearance of popular support, to attack opponents, to flood public space with a particular message. If political actors — regardless of their persuasion — benefit from the existence of bots, or are at least not disadvantaged by them, the incentive to regulate disappears.
Help Net Security, in a recent analysis, observed that bot farms have moved to the centre of modern information warfare, using automated accounts to manipulate public opinion, influence elections, and erode trust in institutions. The question that follows naturally — the one any honest citizen ought to ask — is this: who, precisely, benefits from the fact that nothing is being done?
What Might Be Done: Between Technical Solution and Political Will
The most frequently discussed mechanism for countering bots is real-identity authentication: a system by which social media accounts would be linked to a verifiable identity, at least for participation in public conversations with political consequences — electoral campaigns, policy debates, comment sections attached to news. The model already exists in a handful of jurisdictions: South Korea has experimented with it; China employs it — though in very different contexts, and with objectives that have little to do with protecting democratic discourse.
There are serious arguments against this model, and they deserve to be taken seriously. Online anonymity has protected journalists, dissidents, activists, LGBT individuals in repressive countries, and victims of domestic violence. A system of compulsory real-identity verification could become, in the hands of an authoritarian government, an instrument of surveillance and repression. The tension between "knowing who we are talking to" and "the right to speak freely" is genuine, and admits no simple resolution.
Intermediate solutions do exist, however. Optional identity verification with a visible badge — a "verified account" with actual meaning, not the symbol Twitter once sold for eight dollars a month. Mandatory transparency for digital political advertising. Regulatory limits on large-scale coordinated behaviour. Disclosure requirements for platforms regarding the volume of automated traffic they detect. None of these instruments is sufficient alone, but together they would shift the equation significantly.
What is lacking is not the tools. What is lacking is the political will to use them.
A Personal Note: Why I Chose to Restrict Comments
It is not a costless decision. Open comments mean community. They mean real feedback, the kind of dialogue that makes a website something other than a monologue published into silence. And yet I chose to limit comments to friends and followers, for reasons I believe serious enough to justify the sacrifice.
Even with that restriction, I have encountered automated interventions. Messages that appear human at first glance, but carry the stereotyped structure of generated text — too smooth, too uniform, lacking that small roughness that marks genuine thought. In an unrestricted public space, those interventions would have multiplied several times over.
The decision does not solve the problem. It is, more modestly, a form of personal hygiene in an environment where collective hygiene does not yet exist.
Conclusion: Negotiated Reality and What Is at Stake
We find ourselves at a moment when public reality — what we believe others think, which values dominate, which ideas have consensus, what is acceptable to say aloud — is partly constructed by non-human entities. Not entirely. Not invisibly. But sufficiently to alter the temperature of conversation, to discourage authentic voices, to create the impression of majorities that do not exist.
The greatest risk is not that bots will persuade us to believe false things. It is that they will persuade us we are alone in believing true ones.
If some form of real-identity authentication is not introduced — or at least a body of regulation capable of limiting automated activity at scale — the boundary between human conversation and its simulation will, within a very short time, become impossible for the ordinary user to trace. And a democracy that cannot distinguish the voice of its citizens from the echo of machines is no longer, in any functional sense, a democracy.
The question is not whether we know this. We do. The question is who, precisely, finds the current chaos preferable to order. And why.
Bibliography
-
Imperva / Thales — 2025 Imperva Bad Bot Report: How AI is Supercharging the Bot Threat (April 2025) The principal annual reference report: confirms that automated traffic crossed the 51 per cent threshold in 2024, with malicious bots accounting for 37 per cent of all global traffic. Primary source for quantitative data in Sections I and II. URL: imperva.com/blog/2025-imperva-bad-bot-report
-
Imperva / Thales — 2024 Imperva Bad Bot Report (April 2024) Reference data for 2023: 49.6 per cent bot traffic, the fifth consecutive year of growth in malicious activity. Used for medium-term trend context. URL: businesswire.com/news/home/20240416225637/en
-
Ferrara, Emilio — Bots, Elections, and Social Media: A Brief Overview (USC Information Sciences Institute / arXiv, 2019) Academic analysis of bot operations during the 2016–2018 American elections and the 2017 French presidential election. Empirical foundation for Section III. URL: arxiv.org/pdf/1910.01720
-
Johnson, Neil et al. — PNAS Nexus (2024) Study projecting the use of generative AI in electoral disinformation campaigns, with estimated impact across more than 50 countries. Cited in Scientific American, February 2024.
-
Carley, Kathleen M. — Carnegie Mellon University / Scientific American (February 2024) Researcher cited directly: "It's bots plus generative AI." Context: How AI Bots Could Sabotage 2024 Elections around the World. URL: scientificamerican.com/article/how-ai-bots-could-sabotage-2024-elections-around-the-world/
-
Indiana University Observatory on Social Media — Inauthentic Coordinated Behaviour (The Conversation, 2024) Methodology for detecting coordinated influence operations; limitations of current AI content detectors. Used for Section V. URL: theconversation.com/how-foreign-operations-are-manipulating-social-media-to-influence-your-views-240089
-
SAIS Review — Social Media, Disinformation, and AI: The 2024 U.S. Presidential Campaign (January 2025) Documents troll farm activity in the 2024 electoral context, with emphasis on polarisation amplification irrespective of candidate. URL: saisreview.sais.jhu.edu
-
CETaS / Alan Turing Institute — AI-Enabled Influence Operations: Safeguarding Future Elections (2025) Public policy report: assesses the impact of AI disinformation in 2024; concludes that discourse distortion is demonstrable even where direct effects on voting outcomes remain uncertain. URL: cetas.turing.ac.uk/publications/ai-enabled-influence-operations-safeguarding-future-elections
-
Arkose Labs / Help Net Security — Bot Farms Are Now Running the Global Disinformation War (October 2025) Industry security perspective on bot farm infrastructure as an instrument of information warfare. URL: arkoselabs.com/latest-news/bot-farms-disinformation-war
-
ResearchGate — Social Bots in Election Campaigns: Theoretical, Empirical, and Methodological Implications (2018, updated 2024) Academic synthesis of the literature on political bots, coordinated behaviour, and narrative manipulation mechanisms. URL: researchgate.net/publication/328929907
Author's Note: On Limits, Tensions, and Rhetorical Choices
Any text about bots and digital manipulation lives, inevitably, in the shadow of its own conditions of possibility. I am writing about machines that imitate people — on the internet, the very medium in which that imitation has become a serious problem. It is an irony I have tried not to ignore.
On empirical limits. The statistics cited here — 51 per cent automated traffic, 37 per cent malicious bots, 9–15 per cent fake accounts on social platforms — come from credible sources, but not neutral ones. Imperva is a cybersecurity company with a direct commercial interest in heightening the perceived scale of the threats it addresses. Its reports are methodologically rigorous, but they are not produced by independent academic institutions. I have chosen to cite them with that caveat — not to invalidate them, but to situate them correctly. The exact figures may vary; the trend they describe is confirmed by enough independent sources to be treated as real.
On the unresolved tension around anonymity. This piece implicitly advocates for forms of online identity authentication, but it cannot ignore the fact that anonymity has protected — and continues to protect — vulnerable people. That tension has no simple resolution, and I make no claim to have found one. I chose to name it explicitly in the main text; but it bears repeating here that any public policy moving in this direction would need to be constructed with serious safeguards against the abuse of identity data by states or corporations.
On the rhetorical devices employed. I opened the piece with a threshold — the moment when a number becomes a symptom. It is a framing technique that lends a statistic a moral significance it does not carry on its own. I chose it deliberately, because I believe it is justified; but I name it here in the interest of transparency. The closing rhetorical question — "who, precisely, finds the current chaos preferable to order?" — was deployed as an instrument of suspended certainty, not as an accusation. The question has several plausible answers, each with different implications; my intention was to open the space for reflection, not to foreclose it with a predetermined conclusion.
On comparisons and analogies. The social proof analogy from evolutionary psychology is useful but reductive: the mechanisms by which disinformation actually influences electoral behaviour are more complex and less linear than the analogy implies. I used it as a point of entry, not as a complete explanation. The comparison of restricting comments to "personal hygiene" is an honest metaphor for the limits of individual action in the face of a structural problem — and an acknowledgement that personal solutions do not substitute for collective ones.
On the balance between narrative and documentation. I chose an essayistic register rather than a technical report. That means I have occasionally sacrificed precision for accessibility. Those who wish to go deeper into the technical taxonomy of bots, detection mechanisms, or BaaS architecture will find considerably more substance in the Imperva reports and the academic works cited above. This text is an invitation to conversation, not an encyclopaedia.
One final honesty: I do not know whether this piece will be read by more human beings or by AI crawlers. Statistically, in 2024, the odds were roughly even. I am writing for people, in the hope that they will yet outnumber the machines.
Translator's Note
This essay was originally written in Romanian and adapted into British English for an anglophone readership. The translation sought to preserve not only the meaning but the rhythm and rhetorical texture of the original — the deliberate alternation between short, punchy sentences and longer, more ruminative ones; the suspended questions; the slight melancholy that runs underneath the analytical surface. Where necessary, long Romanian periods were reshaped into more naturally English structures without, I hope, any loss of depth or nuance. The cultural and institutional references (to Romania specifically, to the Romanian political context) have been retained, as they contribute to the authenticity and specific gravity of the argument. British spelling and conventions are used throughout.