How Astroturfing Creates the Appearance of Consensus

Imagine seeing a hundred accounts endorse the same product or repeat the same political argument. Before treating that activity as a hundred independent judgments, ask what those accounts represent. Are they separate people? Are they openly organizing together? Are some invented identities, automated accounts or participants with undisclosed connections? Those questions describe different possibilities, not conclusions about any particular crowd.

Astroturfing is the presentation of organized activity as independent grassroots participation. Keller and colleagues’ research on political astroturfing centers precisely on participants pretending to be ordinary citizens acting independently. The central issue is the representation of independence. Whether the work is performed by software, paid people or a mixture requires additional evidence. Keller et al., Abstract.

Separate the machine from the deception

A social bot uses software to carry out social-media activity. Automation itself does not establish a deceptive purpose. Ferrara and colleagues’ foundational review discusses both harmful bots and useful automated services, including news aggregation and customer responses. An account’s automation, the identity it presents and the relationship behind its activity therefore deserve separate questions. Ferrara et al., opening sections.

Consider a clearly labeled automated notice from a transport service. Its automated delivery does not make its stated identity false. Conversely, imagine a person manually operating an invented customer persona. Human authorship does not make that identity truthful. These hypothetical examples show why calling everything suspicious a bot can obscure the actual allegation.

Use a simple description before choosing a label: this action appears automated; this identity is unverified; this relationship was undisclosed; these accounts show repeated coordination. Keep foreign involvement as another question. None of those descriptions, alone, identifies a foreign sponsor.

Real employees can create misleading endorsements

The FTC’s Sunday Riley case concerned people with an undisclosed commercial relationship. The agency alleged that, between November 2015 and August 2017, company managers posted product reviews on Sephora using fake accounts or asked employees to do so. The complaint also alleged a failure to disclose employment connections. These are the FTC’s allegations about that period. FTC complaint, paragraphs 8–15.

The final order announced in November 2020 prohibited misrepresenting reviewers as independent or ordinary users and required disclosure of unexpected material connections. The respondents neither admitted nor denied the complaint’s allegations, apart from jurisdictional facts. The case illustrates why a review investigation should examine the author’s relationship to the seller as well as the words on the page. FTC decision and order, pp.1, 3–4.

Popularity itself can be the product

In its 2019 Devumi complaint, the FTC alleged the sale of fake followers, subscribers, views and likes. That October, a federal court entered a stipulated order restricting the defendants’ sale of social-media influence indicators and deceptive representations. The settlement did not amount to an admission of the complaint’s allegations. FTC complaint, paragraph 9; court order, pp.1–3.

Here the relevant question is what a displayed number represents. A count may look like evidence of attention or support, but its meaning depends on how the interactions were produced. If you are evaluating credibility, ask for evidence beyond the total: identifiable sources, substantiated claims and independently checkable work. A large number should not do all the reasoning for you.

A public-comment record is not a head count

A 2021 New York attorney general investigation found nearly 18 million fabricated comments among more than 22 million submissions in the FCC’s 2017 net-neutrality proceeding. It identified over 8.5 million fake comments from a broadband-industry-funded campaign opposing net neutrality and more than 9.3 million additional fake comments supporting it. The latter used fictitious identities; the industry-linked effort involved identities of real people used without consent. New York report, pp.3–6 and 27.

A critical qualification appears in the same report: investigators said they had not found evidence that the broadband companies funding and organizing the lead generators had direct knowledge of the fraud, and had not found that those companies violated New York law. Keep that limitation with the funding claim. New York report, p.3.

For assessing participation, distinguish a real person consenting to a template submission from someone’s identity being used without permission. Similar wording is a starting observation; identity and consent are the consequential questions. Nor does the existence of fabricated comments establish how much they changed the final regulatory decision.

Look at relationships, then test alternatives

Keller and colleagues studied a South Korean intelligence-service campaign during the 2012 presidential election using participating accounts listed in court proceedings. They reported that coordination traces distinguished those accounts better than the individual-account characteristics commonly used in bot detection. That is a finding about a documented historical case, not a universal test for every online discussion. Keller et al., Abstract.

A wider methods study by Pacheco and colleagues examined shared images, retweets, hashtag sequences and timing to uncover likely coordination. For readers assessing such research, the important question is what combination of evidence supports the proposed relationship. Ask whether a public organizing event, a shared news source or ordinary group participation could also explain the pattern. Pacheco et al., Abstract.

Repeated activity may justify further investigation without justifying a public accusation. Ask the analyst to distinguish coordination from deception, and both from attribution. Preserve dates and original context. Avoid publishing personal information about ordinary participants merely because their accounts appear in a suspicious cluster.

A bot score needs an explanation

In a 2020 evaluation of Botometer v3, Rauchfleisch and Kaiser examined five datasets in English and German over three months. They found classification problems and scores that varied over time, producing both false positives, where people were classified as bots, and false negatives, where bots were classified as people. These results concern the tested version and setting; they are not a current benchmark of every detector. Rauchfleisch and Kaiser, Abstract and Discussion.

Before relying on a score, ask which model and version generated it, which accounts it was tested on, and how errors were checked in the relevant language and setting. Treat a score as one input to review. Do not convert a threshold into a definitive allegation that someone is fake, paid or controlled by a government.

AI-generated prose introduces another distinction. Yang and Menczer’s 2024 case study identified a network of 1,140 accounts through heuristics and manual review; the accounts appeared to use ChatGPT. The authors reported that coordination was useful for identifying the network, while the text classifiers they tested failed to distinguish its accounts from human accounts in the wild. The result is case-specific and historical. Yang and Menczer, published Abstract.

What a careful reader can do

Start with the claim you actually need to assess. For a product recommendation, examine the reviewer’s disclosed connections and supporting evidence. For a purported public movement, ask what is known about organizers and participation. For an alleged account network, request the documented links and the limits of the investigation. These checks are more precise than deciding whether the prose feels human.

Then separate activity from results. Posting, receiving a view, earning a response and changing someone’s decision are different events. Ask which one a report measured. An impressive account total can document scale without demonstrating persuasion. Leave unknowns visible, including how much activity an investigation might have missed.

Manufactured consensus becomes easier to discuss accurately when the allegation is specific: fabricated participation, concealed employment, misrepresented popularity or deceptive coordination. Use the narrowest description the evidence supports. That protects the integrity of the investigation and the people whose genuine participation should remain part of public debate.