The Artificial Crowd: How Bots Falsify the Social Temperature

Four hundred comments, twenty-four accounts

A small town council announces a plan to restrict car traffic in the centre. Four hundred comments appear beneath its post overnight. ‘Residents are furious,’ an editor says the next morning, commissioning a quick story. The reporter opens the thread. The same turns of phrase recur in slightly different forms; some accounts are new; almost all point to one website. Perhaps a party, company or activist group is behind them. Perhaps automation is involved. Perhaps the opponents are simply unusually well organised.

The example is hypothetical. The editorial mistake is not: converting platform activity into a statement about society. Four hundred comments are four hundred events in a database. We do not yet know how many people produced them, where those people live, whether they acted independently, or what the silent residents think. Yet the headline ‘The internet rejects the plan’ is already warming up in the changing room.

A fake image can mislead us about one event. An artificial crowd can mislead us about what most people supposedly believe. Generative AI lowers the cost of producing plausible comments, adapting personas and operating multiple accounts. Its most consequential effect need not be persuading anybody that one particular message is true. It can be enough to alter the

social temperature: make one position look ordinary and its opponents look lonely.

Real people pay the bill. A local teacher decides not to comment because every supportive voice attracts a wall of abuse. A journalist amplifies an apparent trend. A councillor mistakes manufactured noise for a democratic mandate. A minority loses visibility not because it lost a fair argument, but because somebody placed the thermometer over a radiator.

The spiral of silence: people look before they speak

Elisabeth Noelle-Neumann developed the spiral of silence theory to explain why some opinions become publicly louder while others disappear from view, even when they survive in private conviction. People monitor the climate of opinion. If they perceive their view as growing or dominant, they are more likely to express it; if it appears to be losing and they fear isolation, they are more likely to remain silent (Noelle-Neumann, 1974, pp. 43–51).

The theory does not say that everybody obeys a majority. Topic, commitment, reference group, personality, safety and social norms all matter. It describes a feedback loop. The visible advantage of one position can increase its supporters’ willingness to speak and reduce that of its opponents. The advantage then becomes still more visible. Public expression is not a census of private belief; it is partly a response to what speaking seems likely to cost.

Ewa Nowak-Teter captures the mechanism in her analysis of online public opinion: ‘Our inclination to express our views therefore depends on how we assess the climate of public opinion’ (2023, p. 142, author’s translation). She also discusses fear of isolation and the quasi-statistical sense through which people estimate the distribution of positions. This ‘sense’ is not professional polling. It works with whatever signals the environment makes available.

Those signals once came mainly from conversations, newspapers, television and immediate social reactions. Today, trend lists, likes, repost counts, ranked comments and algorithmic recommendations act as thermometers too. Their precise-looking numbers borrow the authority of measurement. But ‘12.4K’ does not explain what has been measured. It may represent thousands of

citizens, hundreds of repeat contributors, people outside the community, automated accounts, or some mixture of all four.

The spiral of silence therefore receives new fuel in a platform environment: we observe not society itself, but a mediated, ranked image of society. Online public opinion emerges amid anonymity, algorithmic selection and fragmented discussion (Nowak-Teter, 2023, pp. 191–214). This is not a mirror held up to citizens. It is a mirror that selects the reflection, increases the contrast and sells the space beside it.

From bot to climate of opinion

Social bots did not begin with generative AI. Earlier research described accounts that automate activity, imitate people and intervene in platform discussions (Ferrara et al., 2016). Newer systems, however, make variation cheaper. Yesterday’s crude bot repeated a sentence. Today’s model can generate hundreds of stylistically different comments, change tone, answer objections and maintain a persona with an apparently consistent biography.

This is where astroturfing matters: organising a campaign so that it resembles a spontaneous grassroots movement. The name comes from synthetic grass pretending to be the natural thing. It is an unusually exact metaphor for political communication. The operator does not need to manufacture a perfect human being. It need only cover the digital square in an even green surface that looks, from a distance, like living civic energy.

Generative astroturfing can establish a frame before facts arrive, flood replies until conversation becomes exhausting, mass-report selected material, push a slogan into a trending list, or target people who speak against it. Even if no one changes their belief, some will decide that voicing it is not worth the punishment. A synthetic account can produce authentic silence.

Yet the operator is not the only source of power. Platforms decide which reactions become visible and which counts users see. Newsrooms choose whether to describe the result as ‘an online storm’. Politicians can present activity figures as evidence of consent. AI changes the public arena partly through institutions and systems that mediate information, content production and the visibility of interventions (Jungherr & Schroeder, 2023). The artificial

crowd is therefore the product of an entire chain: generation, coordination, ranking, measurement, narration and political use.

That chain matters because responsibility often evaporates inside it. The bot operator blames the platform; the platform describes neutral engagement; the newsroom says it merely reported the trend; the politician cites the newsroom. By the end, everyone has touched the thermometer and nobody admits to moving it.

Three questions that must not be glued together

Discussion of bots often collapses three different questions. First: is the account operated by a real person? Second: is the activity coordinated? Third: does the visible sample represent the population? An answer to one does not settle either of the others.

Real people can join coordinated campaigns. A bot can circulate a genuine document. An authentic, spontaneous group can be extremely vocal and wholly unrepresentative. Conversely, anonymity is not proof of manipulation. It may be a condition of free speech for a whistle-blower, an abuse survivor or a resident of an authoritarian state. A reporter hunting for the magical label ‘bot’ can miss the more important methodological problem.

The same warning applies to synthetic participants in research. A model can simulate many personas quickly and help test questionnaire wording. It is not a sample of society. Replacing human participants with large language models can flatten and misrepresent identity groups (Wang et al., 2025). A model reproduces patterns in its data and design. It does not acquire the lived experience of a person merely because a prompt gives it that person’s label.

The useful editorial question is therefore not only ‘Are these comments real?’ but ‘What, exactly, do these comments allow us to say?’ A comment section may demonstrate that a narrative is circulating. It can reveal argumentative strategies, conflict language or signs of coordination. It will rarely demonstrate what percentage of residents supports a policy. A microscope is excellent for viewing a cell. It remains a poor instrument for weighing an elephant.

Human-centred communication begins with the right to dissent

The ethical stake is freedom of expression without fabricated majority pressure. Freedom of speech is not merely the absence of a formal ban. It includes a meaningful ability to participate in a space where synthetic mass and automated harassment do not make contribution unreasonably costly. If one person must answer a thousand disposable personas, debate is formally open and practically barricaded.

The second stake is the dignity of a minority voice. A person must not be weighed against a pile of accounts as though every unit on a dashboard were an independent citizen. A human participant brings experience, vulnerability, responsibility and the possibility of consequences. A synthetic persona can vanish after a campaign without correcting the record or paying a reputational price.

The third is editorial responsibility. ‘The internet thinks’ is tempting because it saves time on method. The internet thinks nothing. Particular people hold beliefs; accounts perform actions; platforms arrange their visibility; researchers choose what to collect and count. Good journalism keeps those levels visible in its language.

Instead of announcing that Britons are outraged, a report might state that critical comments predominated in the public posts examined, while the sample was not representative and part of the activity showed signs of coordination. The sentence is longer. Democracy sometimes requires more than a three-word headline.

Human-centred moderation also requires restraint. Protecting discussion cannot become an excuse to expose anonymous critics, collect unnecessary personal data or treat unusual language as proof of automation. The aim is not to decide who deserves a voice. It is to prevent manufactured visibility from masquerading as the voice of everyone.

The Opinion Thermometer: Six Checks Before Writing ‘Everyone Is Saying It’

Use the OPINION THERMOMETER whenever a story draws on comments, trends, hashtags, engagement figures or synthetic respondents to describe a social mood. The reporter completes it; an editor or data-literate colleague checks it. Its pass condition is simple: the story must clearly separate platform activity from public opinion.

1. Define the unit. Record whether you counted posts, comments, reactions, accounts, unique users or verified people. One account may publish hundreds of messages; one person may operate several accounts. Pass when the named unit remains the same throughout the story and never quietly turns into ‘people’.

2. Record provenance. Note the collection period, method, available location signals, account age and any declared automation. Do not expose ordinary users or gather personal data merely because it is technically possible. Pass when readers can see the boundaries of the dataset and its major gaps.

3. Test coordination. Look for identical or near-identical phrases, shared links, synchronised timing, abrupt spikes and networks of mutual amplification. These are indicators, not verdicts: genuine activists share materials too. Pass when every claim about coordination carries a confidence level and at least one plausible alternative explanation.

4. Examine distribution. Ask whether activity comes from many independent parts of a network or a small, intense cluster. An average can hide the fact that ten accounts generated half the traffic. Pass when the story reports the concentration among the most active contributors, or states honestly that this cannot be calculated.

5. Test representativeness. Define the population about which you wish to speak. Did its members have a roughly comparable chance of entering the sample? Public comments almost never meet that condition. Pass when the article separates an account of platform discussion from a claim about residents, voters or society.

6. Publish uncertainty. State what remains unknown: the automation rate, the number of silent users, the effect of ranking, gaps in access and possible deleted material. Avoid a single, polished percentage when the foundation is brittle. Pass when a reader can distinguish observation, inference and hypothesis.

If the story cannot pass, change the claim rather than decorating it with a disclaimer. Report the campaign, language, network or signs of coordination. Do not call them the representative will of a population. Uncertainty is not an embarrassment hidden in the technical notes; it is part of the finding.

What the thermometer cannot tell us

No checklist will identify every bot. A sophisticated operation can look organic, while genuine mobilisation can look mechanical. Platform data are incomplete, interfaces change and an outside researcher cannot see every signal. An absence of evidence for coordination is not evidence of its absence.

The OPINION THERMOMETER does not decide whether a position is true. Majorities can be wrong, minorities can manipulate, and an automated account can share accurate information. The method protects against a false claim about the distribution of opinion; it does not adjudicate the underlying dispute.

Nor should all organised communication be treated as suspect. Trade unions, social movements and neighbourhood groups coordinate messages because collective action requires coordination. Communication networks can support both power and counter-power (Castells, 2009). The ethical test concerns deception, automation, fabricated personhood, harassment and claims of representativeness – not the mere fact that people speak together.

Finally, technical detection can itself create injustice. Systems trained to flag ‘inauthentic’ language may misclassify non-native speakers, disabled users, people who dictate messages or communities with unfamiliar idioms. A newsroom should never publish a person’s identity or accuse them of being a bot on a statistical score alone. A responsible conclusion may be ‘we observed coordinated-looking activity’, not ‘these named people are fraudulent’.

Do not fix the thermometer by silencing the patient

The artificial crowd is dangerous because it exploits a democratic ability we genuinely need: watching and listening to other people. We cannot simply stop noticing the climate of opinion. We can learn to distinguish the temperature from the placement of the sensor.

For newsrooms, that means abandoning the lazy phrase ‘the internet thinks’. For platforms, it means meaningful transparency about automation, ranking and coordinated activity. For researchers, it means refusing to treat synthetic personas as a discount substitute for lived experience. For users, it means remembering that the number beside an icon counts events inside a system, not citizens’ consciences.

The practical response is not a machine that declares which voices are human and closes the gate. It is a chain of accountable human judgements: define the unit, inspect provenance, test alternatives, protect vulnerable speakers, qualify the claim and preserve a route for correction. Technology may help to find patterns; it must not receive the authority to decide who counts as a person.

A communication space keeps human beings at its centre when they can speak without fighting an army of masks, and when their silence is not mistaken for non-existence. Before announcing a social storm, then, check whether you are looking through a window or at a wind machine.