Imagine opening a discussion about a new policy. The first comment says it is a disaster. Several replies agree, someone supplies a personal story, and a small crowd applauds. Before reading much about the policy itself, you have already been shown what looks like public opinion.

Now imagine that the first comment and most of the replies came from the same operation. The argument would arrive with its own audience, helpfully assembled in advance.

That is the part of dead internet theory worth spending some time with: the possibility that the apparent crowd is being manufactured, and real people are taking their cues from it.

The claim⌗

A widely circulated account appeared on the Agora Road forum in January 2021, posted by a user called IlluminatiPirate, who credited earlier imageboard contributions.[1] It alleged that large amounts of apparently human content were produced by artificial intelligence (AI) systems and paid influencers hiding who they work for, with the aim of shaping consumers and culture.[1] The post also advanced a much broader allegation of US government-led, AI-powered manipulation of the world’s population.[1]

In this account, artificial content supplies the conversation, disguised participants supply the reaction, and real people gradually follow the resulting trends.[1] Its author pointed to recycled discussions, repeated replies and the feeling that a once-lively internet had become strangely empty.[1] Those observations are the theory’s starting point: a busy-looking place that no longer feels inhabited in quite the same way.[1]

In the forum’s account, control over the social setting is the point.[1] If someone can manufacture the voices around a reader, they can try to influence which ideas appear ordinary, which appear ridiculous and which seem too widely accepted to question.

This is the social-engineering angle: arranging the apparent company a person keeps in order to steer their judgement. The interesting target is the moment a reader thinks, “Perhaps everyone else knows something I don’t.”

What is documented⌗

Investigations have described fabricated identities and staged engagement.[2][6] Experiments have measured how people responded to social signals and AI conversations.[4][5]

EvidenceWhat the record supportsWhat needs its own evidence
Fictitious profilesInvestigators described an operation posing as ordinary individuals.[2]How many real people it persuaded
Staged repliesOpenAI (ChatGPT’s developer) identified campaigns generating reactions to their own material.[6]Genuine popularity outside those networks
Manipulated ratingsA randomised experiment found that an initial vote affected later voting.[4]The effect in a different setting
Fabricated newsroomsAnthropic (Claude’s developer) reported linked sites and commenting accounts.[7]A single controller of the wider internet

How a manufactured crowd could work⌗

Take an imaginary discussion about a town replacing a public park with a shopping development. One person could create the impression of several different local reactions:

  1. Start the story. An account posts that the development is already hugely popular.
  2. Supply the neighbours. Other controlled accounts reply as a parent, a shopkeeper and a long-term resident.
  3. Create apparent agreement. They support one another and treat objections as something the town settled weeks ago.
  4. Invite real participation. A genuine resident joins a discussion whose starting conditions have already been arranged.

The important deception in that example is independence. Three people with different reasons for supporting a plan suggest something different from one person supplying three biographies. A reader may give the apparent agreement more weight because it seems to come from several lives and interests.[4]

The fake crowd could also stage an argument. A clumsy opponent would give its preferred position an easy win. The entire exchange could be theatre, with a real reader arriving halfway through and assuming the cast had bought their own tickets.

Profiles designed to look like neighbours⌗

In July 2024, the US Justice Department announced the seizure of two domain names and searches of 968 social-media accounts associated with an alleged AI-enhanced bot farm.[2] A reproduction of the announcement on Maryland’s government website says the operation created fictitious profiles, often pretending to belong to people in the United States.[2] According to the affidavits described in that announcement, operators used the profiles to promote Russian government objectives.[2]

The announcement also described a joint advisory involving American, Canadian and Dutch authorities, and said X suspended remaining accounts identified in the court documents.[2] Its figures concern domains and accounts searched, rather than an audience survey or a finding about changed votes.[2]

A message presented as a local person’s view arrives with a different implied relationship from a message clearly labelled as a foreign government’s position. Concealing the speaker changes how the message asks to be trusted.

Fake engagement, then fake newsrooms⌗

OpenAI’s May 2024 report described campaigns using generated replies to create apparent engagement with their own posts.[6] It documented one conversation in which the initial post and subsequent comments were all generated through its models.[6] The same report found that operators mixed AI-generated material with manually written text and copied memes.[6]

The relevant unit is the operation: who sets its purpose, supplies its identities and decides what gets posted. A real person operating the machinery can be part of the deception just as much as an automated account.[6]

A September 2026 report from Anthropic, the company behind Claude, described a commercial influence operation using roughly 70 fabricated news websites.[7] It reported matching social accounts and more than 250 inauthentic commenting accounts supporting the network.[7] The operation published at least 8,913 articles in about 20 languages, according to Anthropic’s investigation.[7]

Anthropic said the network shifted its political stance according to paying clients, and reported little observable engagement from real audiences for most of the content it identified.[7] Its investigators found no evidence of government direction in that commercial case.[7] A reader following three links could believe they had found three separate checks on a story, when the links lead back to one production line.[7]

Can the crowd change anyone’s mind?⌗

A 2013 study by researchers Lev Muchnik, Sinan Aral and Sean Taylor tested how previous ratings affected later votes on a social-news website.[4] Across five months, they randomly assigned 101,281 comments to receive an initial positive vote, a negative vote or no intervention.[4] Positive social influence increased the likelihood of a positive rating by a relative 32%, and the researchers found average final ratings 25% higher in the positively manipulated group.[4] The effects depended on the subject and social relationships involved; negative manipulation also produced a different pattern, including corrective voting.[4]

The website gave each comment a running score: up-votes minus down-votes.[4] Readers saw that score alongside the comment, and comments were not arranged in order of popularity.[4] People using the site were unaware of the experiment and could not trace a vote to a particular user.[4] Randomly assigning the first vote let the researchers compare groups whose comments were similar in expected quality, with the early approval or disapproval as the deliberate difference.[4] That design tested the effect of visible approval while trying to hold the quality of the actual comment steady.[4]

A study published in the research journal Nature in December 2025 randomly assigned participants to discussions with an AI model advocating an election candidate in the United States, Canada or Poland.[5] Its abstract reports significant changes in candidate preference, with effects larger than those typically found for traditional video advertisements.[5] The authors also reported that some information supplied by the models was inaccurate.[5]

What is claimed but not established⌗

The claim that most of the internet has been replaced needs a clear answer to “most of what?” Visits, accounts, published words and the conversations a person actually reads are different things to count.

Imperva, a cybersecurity company, reported in its 2026 Bad Bot Report coverage that automated traffic accounted for more than 53% of web traffic in 2025.[3] Its account includes familiar background automation such as search-engine crawlers and scripts, as well as newer AI agents.[3] It also describes bots querying prices and availability, activity that can happen without producing a single public comment.[3]

A hypothetical bot could request a thousand pages while one person writes a thoughtful reply. A traffic count would register plenty of automation; an authorship count would tell a different story.

For dead internet theory’s larger allegation, the missing bridge is coordination. Establishing a coordinated replacement of human participation would require connections between those operations, a measure of their wider prevalence and evidence of the supposed shared direction.

The strongest explanations⌗

Deliberate, coordinated manipulation is a serious explanation for a particular network when investigators can connect its accounts, infrastructure and activity. The government announcement and the platform reports examined here describe networks with identifiable aims and shared operation.[2][6][7] This explanation becomes stronger as the paper trail gets more specific: instructions, funding, shared control and published output that matches the plan.

Commercial influence for hire offers another explanation for apparently political crowds. Anthropic’s commercial case describes an operator changing positions to serve customers.[7] Two contradictory campaigns could come from similar businesses, or even the same business, without belonging to one global political project.

A mixture of people and automation fits OpenAI’s observation that the campaigns it investigated combined generated output with human-written material.[6] That makes “human or bot?” a less useful question than “who controls this activity?” Paying a person to pose as an independent customer and generating a fictional customer can create the same misleading impression of independence.

OpenAI’s May 2024 assessment found that the campaigns it covered had not meaningfully increased their reach or engagement through its services.[6] Anthropic’s later commercial example likewise described large output with limited authentic engagement.[7]

What remains unknown⌗

The sources here do not provide a representative census of human-written conversation across the whole internet. They also leave commissioning gaps: Anthropic could not independently confirm which customers commissioned the commercial operation it described.[7]

Durable influence is another open question. To assess it, evidence would need to follow real people beyond exposure and measure what changed: a belief, a purchase, participation in a discussion or an action elsewhere.

Dadbot verdict⌗

The social-engineering angle is the most useful part of dead internet theory because it identifies something specific to investigate: counterfeit independence. A statement backed by ten apparently unrelated people asks for a different level of trust from a statement repeated by one organisation through ten accounts.

Fictitious profiles, self-generated engagement and fabricated newsrooms are all present in the records examined here.[2][6][7] The broader theory would become more convincing with verifiable records connecting those operations to a shared programme, alongside representative measurements showing how much real conversation had been displaced.

For an individual thread, start with the origin of the claim and the relationship between the voices repeating it. Five accounts pointing to one report still leave one report to examine. Separate sources should bring something of their own: another witness, another document, another piece of work.

The question worth keeping is simple: who has put this conversation in front of you, and why does it look as though so many people agree?

This article was written with AI. Which is either a useful disclosure or the dead internet theory doing its own publicity.

Sources and caveats⌗

The government account is an announcement of allegations, reproduced by a Maryland government agency. OpenAI and Anthropic report on activity visible through their services. The Nature study is cited from its accessible abstract. The rating experiment measured behaviour on one site. Sources checked in October 2026.