Category: Manufactured Consensus: Astroturf, Sockpuppets and Reputation Engineering

A study of the machinery used to make popularity, agreement, outrage, and trust appear more organic than they really are.

This category covers astroturf campaigns, sockpuppet accounts, coordinated posting, fake reviews, purchased engagement, reputation management, synthetic communities, brigading, influencer seeding, and other techniques used to manufacture the appearance of public opinion.

Posted on by

Auditing an online reputation without treating disagreement as proof of manipulation

The fastest way to ruin an investigation is to begin with this premise:

Everybody who disagrees with me is fake.

That theory explains everything.

Which is exactly why it explains nothing.

A useful reputation audit starts by separating ordinary negative opinion from evidence of manipulation.

People can dislike the same company, creator, product, movement, or website for entirely real reasons.

Consensus can be genuine.

So can disagreement.

Start with the visible record

Before looking for hidden coordination, document what actually exists.

Collect dates, URLs, review counts, account histories, rating changes, archived pages, public disclosures, moderation actions, and examples of the language being used.

Then classify the evidence.

Is the complaint based on firsthand experience?

Is the reviewer connected to a competitor?

Did several accounts appear at once?

Are the same unusual phrases repeated?

Is there evidence of payment, shared control, or a campaign instruction?

This creates a factual map before the arrows start appearing on the corkboard.

Compare manipulation hypotheses with boring alternatives

Suppose a business receives fifty negative reviews in two days.

Possible explanation one: a competitor organized a review attack.

Possible explanation two: a viral video sent thousands of genuine angry customers to the review page.

Possible explanation three: the business had a real service failure affecting many people at once.

The timing pattern is the same.

The causal story is not.

A rigorous audit tests competing explanations instead of treating the most suspicious one as the default.

Relationships matter more than vibes

The FTC’s current review guidance focuses on concrete distortions such as fake reviews, undisclosed insider relationships, conditioned incentives, suppression of negative reviews, and manipulation that changes the picture consumers receive. See the FTC’s Endorsements, Influencers, and Reviews guidance.

That is a useful template for reputation research generally.

Look for relationships and actions that can be documented.

Do not substitute tone analysis for provenance.

A sarcastic review is not a bot signature.

Three people using the same cliché are not automatically sockpuppets.

A critic who posts often may simply be a critic who posts often.

Report confidence instead of pretending certainty

An audit can classify findings in layers:

  • documented — supported by primary records or platform findings;
  • strongly indicated — several independent signals agree, but direct attribution is incomplete;
  • possible — a pattern deserves further investigation;
  • unsupported — the available evidence does not distinguish manipulation from ordinary behavior.

That vocabulary is less exciting than declaring a giant secret operation.

It is much harder to abuse.

The final test should work against your own theory

Ask what evidence would make you abandon the manipulation hypothesis.

If the answer is nothing, the audit has become a belief system.

Manufactured Consensus is real because hidden coordination, fake identities, paid reputation, and astroturfing are documented practices.

That reality does not grant permission to call every inconvenient crowd fake.

A good audit protects two things at once:

It takes manipulation seriously.

And it leaves real humans the right to genuinely disagree.

Posted on by

Correction campaigns that struggle to undo an established false consensus

The correction rarely gets the same opening weekend as the mistake.

A false claim can arrive first, spread widely, accumulate screenshots, become a joke, enter arguments, and get repeated by people who never saw its original source.

Then the correction appears.

Different day. Smaller audience. Worse headline.

That does not mean corrections are useless.

It means the comparison is structurally unfair.

Early impressions can persist

Psychologists use the term continued influence effect for the finding that misinformation can continue affecting reasoning even after people have received a correction.

A meta-analysis by Nathan Walter and Riva Tukachinsky covering 32 studies found that corrections reduced misinformation’s influence but did not always eliminate it completely. See the published abstract and study.

Later reviews reach a similarly important but less dramatic conclusion: factual corrections generally improve belief accuracy, and evidence for a broad “backfire effect”—where corrections routinely make people more wrong—is weak. See the 2024 review Factual corrections: Concerns and current evidence.

So the useful statement is not corrections fail.

It is corrections do not automatically reset the information environment to zero.

Audience overlap is a separate problem

Suppose a misleading post reaches two million people.

A careful correction reaches 80,000.

The correction could be extremely persuasive among the people who see it and still leave most of the original audience untouched.

The 2024 review notes that people exposed to misinformation may be unlikely to encounter the relevant correction at all.

That turns distribution into part of the correction problem.

A false consensus can therefore persist not only because people reject corrections, but because many people never receive them.

Corrections need their own measurement

Researchers should separate at least three questions:

Did the correction reach the original audience?

Did recipients update their factual belief?

Did that change affect later attitudes or behavior?

Those are not the same outcome.

A platform adding a label may increase factual accuracy without changing a person’s opinion of the larger issue. A corrected statistic may disappear from later discussion while the emotional impression it helped create remains.

Conversely, a strong correction can work very well when it reaches people with clear evidence and an alternative explanation.

False consensus has a memory

Imagine a company is falsely accused of receiving thousands of identical complaints.

The fabricated screenshots spread for three days.

The platform later proves the accounts were coordinated and removes them.

Anyone who sees that evidence can revise their view.

But copies may remain elsewhere, search results may retain old headlines, and people may continue saying, “Didn’t everybody hate that company last year?”

The correction has to chase an impression that already escaped its original container.

Manufactured Consensus is therefore not only about creating a crowd.

It is about the residue that crowd can leave behind.

Corrections matter.

They just have to run uphill after the first story already learned the route.

Posted on by

How authentic participants become amplifiers of a concealed campaign

A real person can spread material from a fake persona.

That does not make the real person fake.

It makes provenance complicated.

Influence operations often attract attention because of their controlled accounts: fabricated identities, front organizations, fake news pages, or coordinated profiles. But material does not necessarily remain inside the network that created it.

Once an ordinary user sees the post and shares it, a concealed campaign can acquire an entirely authentic second layer of distribution.

The amplifier may have no connection to the organizer

Imagine a fabricated local-news account publishes a dramatic story.

A real resident sees it, believes it, and sends it to ten friends.

One of those friends posts it in a neighborhood group. Another journalist quotes it before discovering the source is unreliable.

The later participants may be acting completely independently.

Their accounts are real.

Their reactions are real.

What remains concealed is the origin of the material they are amplifying.

That distinction matters because otherwise an investigation can make an unfair leap: this person shared campaign material, therefore this person was part of the campaign.

The evidence does not support that conclusion automatically.

Influence operations often try to cross this boundary

Meta’s adversarial-threat reporting describes covert networks that created fictitious media organizations, NGOs, and other identities across multiple services while attempting to reach real audiences. In its Q1 2023 report, Meta said it removed the majority of several detected operations before they were able to build authentic audiences. See Meta’s Q1 2023 Adversarial Threat Report.

That phrase—authentic audiences—captures the transition.

A deceptive operation does not need every person in its distribution chain to be deceptive.

It only needs enough real people to pick up the material.

Authentic sharing can erase visible provenance

After several rounds of reposting, the original source may disappear from view.

A user screenshots a claim instead of linking it.

Another rewrites it.

A third posts the statistic without remembering where it came from.

Eventually the message can look like ordinary peer-to-peer conversation even though the first push came from a concealed campaign.

This is one reason network analysis becomes difficult. The boundary between organized distribution and organic distribution is porous.

Investigators should separate roles

A useful analysis distinguishes at least three groups:

  • operators, who created or controlled the campaign;
  • coordinated participants, who knowingly helped distribute it;
  • authentic amplifiers, who encountered the material and shared it independently.

One person can move between categories, but the classification needs evidence.

Payment records, internal chats, shared account control, or campaign instructions can support claims of participation.

A retweet does not.

Manufactured Consensus becomes powerful when fake origin and real enthusiasm mix together.

The final crowd may contain thousands of genuine humans.

That does not make the original source transparent.

And the hidden source does not make every human in the crowd a conspirator.

Posted on by

The difference between campaign reach and persuasive effect

A million impressions is not a million changed minds.

That sounds obvious.

Online influence research forgets it surprisingly often.

Platforms can measure views, impressions, followers, shares, clicks, and estimated reach. Those numbers describe exposure.

Persuasion is a different claim.

To show persuasion, a researcher needs evidence that attitudes, beliefs, choices, or behavior changed because of the campaign.

Reach is easier to count

Suppose a covert network produces 40,000 posts that appear in front of five million accounts.

Those are substantial distribution numbers.

They can tell us that the operation existed, produced material at scale, and reached an audience.

They cannot tell us how many people believed it.

Some users may have ignored it.

Some may have already agreed.

Some may have mocked it, blocked it, or forgotten it five seconds later.

A share can even come from a critic.

The counter still goes up.

A documented example shows the gap

A 2023 Nature Communications study linked longitudinal survey responses from U.S. Twitter users with their exposure to accounts associated with Russia’s Internet Research Agency during the 2016 U.S. election.

The researchers found that exposure was highly concentrated: one percent of users accounted for 70 percent of exposures. They also found that IRA content was outweighed by domestic political and news content in participants’ feeds. Most importantly, across the outcomes they examined, they found no evidence of a meaningful relationship between exposure to the Russian influence campaign and changes in attitudes, polarization, or voting behavior. See the study in Nature Communications.

That result does not prove the campaign had zero effect on every person.

It shows why reach statistics alone cannot establish persuasive effect.

Persuasion requires a different research design

Useful evidence can include:

  • randomized experiments,
  • before-and-after surveys,
  • longitudinal panel studies,
  • behavioral data tied to exposure,
  • credible comparison groups,
  • natural experiments with carefully stated limitations.

Even then, causation can be difficult because people choose what they read and whom they follow.

A campaign may disproportionately reach people already sympathetic to its message. High engagement can therefore reflect selection rather than persuasion.

Big numbers make seductive headlines

“Campaign reached 20 million users” is concrete.

“Campaign changed an unknown number of opinions under conditions we cannot fully identify” is less exciting.

The second sentence may be more accurate.

Manufactured Consensus research should therefore keep at least three measurements separate:

How much material was produced?

How many people encountered it?

What changed because they encountered it?

Those questions require different evidence.

A campaign can fail to persuade and still be worth studying.

It can waste attention, distort a discussion, create confusion, provoke journalists, or simply demonstrate an attempt at manipulation.

But an attempt to influence is not proof of successful influence.

Reach tells us how far the message traveled.

Persuasion asks what happened when it arrived.

Posted on by

Platform takedown reports as evidence with incomplete coverage

A platform takedown report is powerful evidence.

It is also a spotlight, not a census.

When a company such as Meta publishes details about a coordinated inauthentic network it removed, researchers get something unusually valuable: linked accounts, attributed operators, tactics, geographic targets, fake entities, and sometimes cross-platform behavior.

That is far stronger than staring at suspicious posts and guessing.

The mistake is assuming the published cases represent everything that exists.

Takedown reports show detected operations

Meta’s adversarial-threat reports describe networks the company investigated and removed for behavior such as coordinated inauthentic activity, cyber espionage, or coordinated abusive reporting.

In its Q1 2023 report, Meta described six coordinated inauthentic behavior networks and noted that some ran fictitious media organizations, NGOs, and other entities across multiple platforms. See Meta’s Q1 2023 Adversarial Threat Report.

That is useful evidence about those specific networks.

Meta also states in the report itself that its public threat reporting is not meant to reflect the entirety of its security enforcements. The published material emphasizes notable investigations and trends. See the full Q1 2023 report.

That sentence matters enormously.

Detection creates a selection effect

A campaign that gets caught is observable.

A campaign that is never detected is missing from the dataset.

A campaign that is detected but not publicly described may also be missing.

Platforms differ in what they investigate, what data they possess, which policies they use, how much detail they publish, and how often they release reports.

So counts such as twelve networks removed this year cannot safely be translated into there were twelve networks this year.

The first number is an enforcement statistic.

The second is a prevalence claim.

They are not interchangeable.

Reports are strongest for tactics and attribution

Researchers can use takedown reports to study:

  • fake personas and fictitious organizations,
  • shared account administration,
  • cross-platform expansion,
  • purchased engagement,
  • coordinated reporting,
  • geographic targeting,
  • links to companies, governments, PR firms, or other operators when the platform documents them.

They are much weaker for estimating how much manipulation remains undetected.

A 2026 IISS analysis compiled hundreds of disclosed takedowns across Meta, Google, and TikTok and used them to study persistent patterns in inauthentic activity. That kind of aggregation can reveal recurring tactics, but it still begins from platform disclosures rather than an omniscient view of the internet. See IISS’s analysis of platform disclosures.

The absence of a takedown proves very little

If a platform never reports a campaign, several explanations remain possible.

The campaign may not exist.

It may exist but remain undetected.

It may have been detected under a different policy.

It may have been handled quietly.

Or it may operate mostly somewhere else.

Manufactured Consensus research needs platform reports because they provide rare, documented cases with internal evidence.

It also needs humility about the dark space around them.

A takedown report tells us what the platform found.

It does not tell us everything that was there.

Posted on by

Deleted disclosures and the changing visibility of sponsorship

A disclosure can disappear without the relationship ever having been imaginary.

Web pages change.

Bios are rewritten. Sponsor pages are replaced. Old campaign sites expire. A disclosure that sat beside a recommendation in 2022 may be gone from the live page in 2026.

If a researcher looks only at today’s version, the history can flatten into something misleading.

Sponsorship is partly a time question

The Federal Trade Commission’s endorsement guidance says material connections between endorsers and marketers should be disclosed clearly and conspicuously when the relationship would affect how people evaluate the endorsement. See the FTC’s Endorsement Guides.

That makes the placement and visibility of a disclosure important.

But a later investigation may encounter a problem: the original page no longer looks the way it did when the audience saw it.

A creator may have changed the caption. A company may have redesigned the site. A sponsor list may have moved. An old “About” page may now return 404.

The live web is not a reliable historical record of itself.

Archives can preserve an earlier state

The Internet Archive’s Wayback Machine stores dated captures of public web pages. Its help documentation notes that saved pages can continue to exist after the original page changes or disappears. See Save Pages in the Wayback Machine.

That makes archives useful for questions such as:

  • Did a sponsorship label appear beside the original post?
  • Was a funder listed on an older version of the organization page?
  • Did a site’s ownership statement change?
  • Was an affiliate disclosure later removed or relocated?

A dated capture can establish what one archived version displayed at one point in time.

It cannot automatically establish what every visitor saw.

Web archives have holes

Wayback documentation is explicit that some pages are never captured, dynamic elements may fail, images can be missing, and archived pages can be incomplete. See Using the Wayback Machine.

So the absence of a disclosure from one archived snapshot does not prove no disclosure existed anywhere else.

Researchers should compare multiple captures, inspect surrounding pages, save timestamps, and preserve the original URL.

Screenshots can help too, but they need provenance of their own.

Deletion changes visibility, not necessarily history

Suppose an influencer disclosed a sponsorship clearly in the original post, then years later removed the disclosure while cleaning up an old profile.

That is a different historical claim from never having disclosed it.

Suppose instead an archived copy shows the disclosure was added only after public criticism.

That is also different.

The date sequence matters.

Manufactured Consensus research often asks who was behind a message.

Sometimes the answer was once printed in plain sight.

Then the web changed clothes.

Posted on by

Funding trails behind apparently independent online advocacy

Follow the money is good advice.

It is not a magic spell.

Funding records can reveal that an apparently independent organization receives money from a company, foundation, trade group, union, political organization, or other interested sponsor. That relationship may be extremely important.

It still does not tell you everything about who wrote every sentence or controlled every decision.

Public records can expose relationships

For U.S. tax-exempt organizations, Form 990 filings can reveal information about revenue, major expenses, grants, officers, related organizations, contractors, and other parts of an organization’s structure.

But the records have important limits. The IRS says contributor names and addresses on Schedule B generally are not required to be publicly disclosed for most organizations filing Form 990 or 990-EZ. Different rules apply to private foundations and certain section 527 political organizations. See the IRS guidance on public disclosure of contributor identities.

So a researcher may be able to see that an organization received substantial contributions without being able to identify every donor from the public filing alone.

Other useful records can include corporate registrations, grant databases, lobbying filings, contracts, board biographies, archived sponsor pages, annual reports, and disclosures from the funder itself.

Funding establishes a relationship, not automatically control

Suppose an advocacy site receives a $250,000 grant from an industry foundation.

That fact matters.

It does not automatically prove the funder selected the site’s conclusions, approved its articles, or dictated its strategy.

The stronger case for sponsor control needs stronger evidence: contractual language, internal correspondence, shared staff, approval rights, campaign instructions, governance ties, or documented coordination.

The opposite mistake is also possible. A site can describe itself as independent while leaving out a financial relationship that a reasonable reader would consider important.

In commercial endorsement settings, the FTC uses the idea of a material connection: a relationship that could affect how an audience evaluates an endorsement should be disclosed clearly and conspicuously. See the FTC’s Endorsement Guides.

Advocacy funding is not governed by exactly the same rules in every setting, but the interpretive principle is useful.

Hidden relationships change how evidence is weighed.

Motive is the hardest claim

A payment can establish that money moved.

It cannot read anybody’s mind.

A donor may fund an organization because it already shares the donor’s views. An organization may accept money and retain substantial independence. A sponsor may also exert direct control.

Those possibilities require different evidence.

A rigorous funding investigation therefore reports layers separately:

Who paid whom?

What formal relationship existed?

What evidence shows influence over the message?

What remains unknown?

Manufactured Consensus becomes visible when apparently independent voices are connected by relationships the audience was not shown.

The money trail can reveal the wiring.

It should not be asked to prove more than the records actually say.

Posted on by

Network timing as evidence of organized amplification

A hundred accounts posting within forty seconds looks organized.

Sometimes it is.

Sometimes Beyoncé just dropped something.

Timing is useful evidence because coordination often leaves a clock-shaped footprint. Accounts publish the same link in bursts. Replies arrive seconds apart. A cluster repeatedly activates together and then goes quiet together. One account posts and a familiar set immediately amplifies it.

Those patterns are worth investigating.

They are not self-interpreting.

Coordination often creates synchronized activity

Platforms investigating influence operations routinely examine behavior across networks rather than judging one post in isolation.

Meta describes coordinated inauthentic behavior as coordinated efforts using deceptive identities or other inauthentic behavior to mislead people about who is behind an operation. Its threat reports describe networks acting across multiple services and using coordinated account structures. See Meta’s Coordinated Inauthentic Behavior archive.

Timing can help reveal those structures.

If thirty accounts repeatedly share obscure material within the same tiny windows, activate in the same sequence, and interact with the same core accounts, the timing pattern may support a larger network analysis.

But the clock does not tell you why the pattern exists.

Legitimate events also create bursts

Breaking news creates synchronized posting.

So do sports scores, product launches, livestreams, scheduled press conferences, television premieres, earnings reports, game releases, emergency alerts, and fandom campaigns.

A public advocacy group can also tell real supporters, openly, post at noon.

That is organized amplification.

It is not necessarily deceptive amplification.

The same caution applies to automation. Scheduled social-media software can make perfectly legitimate accounts publish at identical times. DIT-145 dealt with timing as a clue to automation; here the narrower question is whether a network of humans or accounts is deliberately amplifying something together.

Those are different claims.

Timing gets stronger when other evidence agrees

Useful corroboration can include:

  • shared administrators or login infrastructure,
  • repeated coordination across many unrelated posts,
  • common campaign links or tracking codes,
  • internal chat instructions,
  • account creation clusters,
  • identical source material,
  • financial or organizational relationships,
  • platform takedown findings.

One synchronized burst after a major event is weak evidence.

A recurring pattern across months, combined with shared control and hidden sponsorship, is much stronger.

This matters because Manufactured Consensus research can become numerology very quickly.

Humans love clocks and clusters.

Give us timestamps and we can draw sinister arrows between breakfast and lunch.

The useful question is not merely did these accounts move together?

It is what evidence explains why they moved together, and was that relationship hidden from the audience?

Posted on by

Shared language as evidence of coordination and its limits

When fifty accounts use the same strange sentence, somebody should probably look closer.

That does not mean the case is solved.

Shared language is one of the easiest signals of possible coordination to notice online. Accounts repeat the same slogan, cite the same unusual statistic, make the same spelling mistake, or arrange several arguments in the same order.

Sometimes that happens because one organizer supplied the language.

Sometimes everybody copied the same article.

Those are not the same finding.

Similarity is a clue about origin

Suppose twenty comments all say:

This proposal is a bureaucratic hammer searching for a digital nail.

That phrase is unusual enough to justify asking where it came from.

If investigators later find a campaign memo instructing participants to use that exact sentence, the similarity becomes useful evidence of a common source.

But repeated wording alone cannot identify the organizer.

People quote press releases. Fans repeat catchphrases. Journalists paraphrase wire stories. Activists share sample letters publicly. Customers copy one another’s troubleshooting answers. Thousands of people can honestly repeat the same sentence after reading it in the same place.

The FCC fake-comment investigation shows the difference

A strong example comes from the New York Attorney General’s investigation into comments filed during the FCC’s 2017 net-neutrality proceeding.

The office reported in 2021 that nearly 18 million of the more than 22 million comments received by the FCC were fake. More than 8.5 million comments impersonated real people in campaigns connected to broadband-industry funding, while another 9.3 million fake comments supporting net neutrality used fictitious identities, mostly submitted by one individual using automation. See the New York Attorney General’s investigation summary.

The important methodological point is that the investigation did not stop at noticing repeated sentences.

It used records from lead generators and campaign operators to establish where the submissions came from and whether the named people had actually participated.

That is a much stronger evidentiary chain.

Language analysis works best when paired with provenance

Useful corroboration can include:

  • shared documents or campaign instructions,
  • identical links with tracking parameters,
  • common account administrators,
  • payment records,
  • synchronized posting tied to a known coordination channel,
  • admissions from participants,
  • platform enforcement findings.

Language similarity becomes more persuasive when several independent signals point toward the same explanation.

It becomes weaker when the wording is generic, widely quoted, or attached to a major public event that gave everybody the same source material.

Manufactured Consensus is easy to imagine because humans naturally notice patterns.

The research problem is harder.

Same words can mean same organizer.

They can also mean same newspaper article.

The responsible investigator does not confuse the first clue with the final answer.

Posted on by

Distinguishing grassroots mobilization from deceptive astroturfing

Grassroots does not mean disorganized.

Real people coordinate.

They form committees, share sample letters, choose hashtags, raise money, build mailing lists, hire staff, schedule demonstrations, distribute graphics, call journalists, train volunteers, and remind everyone what time to show up.

None of that turns a movement into astroturf.

The important question is whether the visible public support represents what it claims to represent.

Coordination and deception are different variables

A useful everyday definition of astroturfing is communication that appears to come from ordinary independent members of the public but actually comes from an organization or interested sponsor in a way designed to make support look broader or more spontaneous than it is. Cambridge’s current definition captures that emphasis on disguised origin. See Cambridge Dictionary’s definition of astroturfing.

Meta uses a similar behavior-first distinction in its coordinated-inauthentic-behavior work. Its enforcement reports focus on networks that coordinate while using fake accounts or deceptive identities to mislead people about who is behind the activity. Meta explicitly says it evaluates the deceptive behavior rather than whether it agrees with the content. See Meta’s explanation of coordinated inauthentic behavior.

That is a useful research discipline.

Do not begin with whether you like the message.

Begin with who is actually speaking.

A real campaign can use identical scripts

Suppose an environmental organization publicly asks 10,000 members to email lawmakers using the same template.

Suppose an industry association does the same thing.

Both are organized.

If the senders are real people who knowingly chose to participate, the repeated language by itself does not make either campaign fake.

Now consider the New York Attorney General’s investigation into comments submitted during the FCC’s 2017 net-neutrality proceeding. The office reported that millions of comments were fabricated, including comments submitted using identities of people who had not agreed to participate. The investigation relied on lead-generator records and other evidence connecting the submissions to paid campaigns. See the New York Attorney General’s 2021 investigation summary.

That is qualitatively different from supporters choosing to send a form letter.

The constituency itself was misrepresented.

A fair classification needs several kinds of evidence

Useful evidence includes:

  • disclosed and undisclosed funding,
  • who created campaign infrastructure,
  • whether participants are real,
  • whether they knowingly joined,
  • whether identities were forged or impersonated,
  • whether accounts share hidden control,
  • whether organizers instructed participants to conceal sponsorship,
  • whether public claims about independence match internal records.

Shared language, synchronized timing, and sudden popularity can support an investigation.

They are weak foundations for a conclusion on their own.

The standard should work in both directions

This matters because “astroturf” is an attractive insult.

It can be used to dismiss genuine public organizing simply because participants are coordinated, professionally supported, or politically inconvenient.

The opposite error is just as bad: treating manufactured identities as authentic public sentiment because real organizations or real humans appear somewhere in the chain.

A rigorous Dead Internet Theory study needs to survive both temptations.

The web contains grassroots movements.

It contains professional advocacy.

It contains paid campaigns.

It contains fake people pretending to be real movements.

The job is not to decide which crowd deserves to exist.

The job is to determine whether the crowd is the crowd it claims to be.