Dead Internet Theory: The Conspiracy That Came True

July 9, 2026

A research position. Measured, sourced, present-all-sides. The subject is a conspiracy theory, and the discipline is the one the whole corpus runs on: separate the documented fact from the myth, trace where the myth came from, and grade every fringe claim by the evidence rather than by who said it. Peer-reviewed and conference work is marked; a vendor with a commercial stake is flagged as such; where a number is contested, the range travels with it. Detector-based figures carry false-positive rates their vendors do not fully characterize, and are cited as detector estimates, not as counts.

There is a conspiracy theory about the internet, and the embarrassing thing about it is that it was partly right. Not right in the way it wanted to be — there is no hidden hand, no gaslighting program, no staged civilizational collapse with the servers left running as camouflage. It was wrong about all of that. But it made one falsifiable prediction, and somewhere between 2023 and 2026 that prediction stopped being paranoid and started being an industry statistic. This is the clean case of the information commons becoming a capture surface — not by conspiracy, but by ordinary incentives running at machine scale. The machine did not need a controller. It needed an ad-tech payout and a cheap generator.

The claim

Dead Internet Theory holds that the internet “died” around 2016 or 2017, and that most of what you now see — posts, replies, traffic, “people” — is bots, AI-generated filler, and coordinated astroturf, with real humans a shrinking minority. In its strong form it is a paranoid metaphysics: a deliberate government-and-corporate psyop to gaslight the population, the humans mostly gone, the lights kept on as a disguise.

That strong form is unsupported. There is no evidence of a single coordinating hand, and the theory’s own believers reach for aliens and collapse to explain the vibe. But the theory made a measurable claim — that the commons would fill with non-human content and non-human traffic until the human share was the minority — and that is the part worth taking seriously, because it turned out to be partly testable, and it partly held.

The tinfoil version was wrong about who and increasingly right about what. That inversion is the whole reason this belongs in the corpus. The web became a captured commons anyway — which is worse than a conspiracy, because there is no one to depose and nothing to leak. You cannot leak your way out of an equilibrium. This is the same emergent-capture pattern the corpus documents everywhere else, arriving at the information environment itself (see capture-mechanism-universal and the running definitions in the glossary).

Where the theory came from

The theory was crystallized in a post titled “Dead Internet Theory: Most of the Internet is Fake,” published January 5, 2021 by a user going by IlluminatiPirate on Agora Road’s “Macintosh Cafe” esoteric board. The post did not invent the frame so much as assemble and name ideas that had already circulated on the forum and adjacent imageboards; earlier threads on Wizardchan and in 4chan’s late-2010s culture had been carrying the same core intuition — that the web felt increasingly automated, curated, and empty of organic human presence (Wikipedia). IlluminatiPirate is a pseudonymous handle; the coinage is attributed to the post, and nothing is asserted about the person behind it.

The theory reached a mainstream audience through Kaitlyn Tiffany’s Atlantic piece, “Maybe You Missed It, but the Internet ‘Died’ Five Years Ago” (2021), which called the post the theory’s “ur-text” and treated the belief as a genuine structure of feeling — “wrong but feels true” — rather than merely mocking it (The Atlantic). A reputable outlet engaging the idea, rather than laughing it off, is what gave it lift.

It is worth holding the two versions apart, because they have aged very differently. The observable core — that the web is dominated by automated systems and algorithmic curation rather than organic human activity, with the “time of death” placed around 2015 to 2016 — is the falsifiable part. The weak version — that elites and corporations deliberately deploy bots to shape discourse — is partially documented as a practice, though not as a total explanation. The strong version is the conspiracy proper: the ur-text’s own formulation, as quoted by The Atlantic, is that “the U.S. government is engaging in an artificial intelligence-powered gaslighting of the entire world population,” with the most extreme variant holding that society already collapsed and unspecified entities keep the internet running as a disguise. No evidence supports either; they are recorded here strictly as the conspiracy’s content, not adopted as a finding.

What came true

The theory predicted a web where non-human content and non-human traffic crowd out humans. Here is the sourced state of that prediction. Every figure below measures a different quantity with a different method; they do not add up to one number, and anyone offering a single “X% of the internet is fake” is reporting their favorite, not the truth. This is the same denominator discipline the model-collapse dossier enforces on the content-share ranges.

Bots overtook humans on traffic. Per the 2025 Imperva (Thales) Bad Bot Report, automated traffic reached 51% of all web traffic in 2024 — the first time in a decade that bots surpassed human activity — with bad bots rising to 37% of all traffic, up from 32% in 2023. Imperva attributes the surge to large language models lowering the cost of building and scaling bots (Thales; Imperva). The same report breaks out AI-crawler activity, naming ByteSpider (ByteDance) as responsible for the largest share of AI-enabled bot attacks at roughly 54%, followed by AppleBot, ClaudeBot, and the ChatGPT user-agent (Thales blog) — an attributed vendor figure for one axis, attack traffic, and not a claim about intent. Keep the caveat: traffic share is request volume, a different axis from content. One human can trigger a thousand bot requests; a bot can serve human-written text. “51% bots” leaves 49% human traffic — not “humans are 12% of the web.”

AI content is flooding the commons. NewsGuard’s AI tracking documents the content-farm explosion: it began identifying Unreliable AI-Generated News sites in May 2023 with 49 sites; by February 2024 the count exceeded 700; the current tracker lists 3,749 such sites across 16 languages, defined by four criteria including clear evidence of AI production, publication without significant human oversight, and no disclosure (NewsGuard; NewsGuard). On Facebook, a Stanford Internet Observatory / Georgetown CSET study (DiResta and Goldstein, 2024) examined 120-plus pages each posting AI-generated images — the “Shrimp Jesus” phenomenon — that drew hundreds of millions of interactions, with Facebook’s Feed recommending unlabeled AI images to users who did not follow the pages; the documented motivations were profit and clout, not ideology (HKS Misinformation Review; arXiv; 404 Media; Forbes). The distribution side of this — the recommender actively boosting the slop — is its own subject; see algorithmic amplification.

At the platform-text level the numbers come from detectors, and are cited as such. WIRED commissioned detection firms to sample Medium: Pangram Labs analyzed 274,466 posts over six weeks and estimated over 47% were likely AI-generated, and Originality.ai found Medium’s likely-AI share rose from about 3.4% in 2018 to over 40% in a 2024 sample. Originality.ai separately estimates that over half of long-form LinkedIn posts are likely AI-generated post-ChatGPT, and about 15% of Reddit posts in 2025 (Originality.ai; Originality.ai). These carry real false-positive rates. And the book layer went the same way: Amazon’s self-publishing platform was flooded after ChatGPT, prompting Amazon to limit authors to three self-published titles per day in September 2023 and to require AI disclosure that December — measures widely reported as ineffective against the volume (NPR; Rolling Stone).

Bots that post like people. The astroturf prediction — the weak version — is now documented and automated. “Anatomy of an AI-powered malicious social botnet” identified “fox8,” a network of roughly 1,140 X accounts apparently using ChatGPT to generate human-like posts and replies; notably, AI-text detectors including OpenAI’s own and GPTZero failed to reliably flag the botnet’s output, and the authors warn it is “the tip of the iceberg” (Yang and Menczer, arXiv; Tech Policy Press). This is astroturf with the human removed — the troll’s craft replaced by a token generator, the marginal cost driven to zero. The human version of that craft — the scam economies and griefing operations people built by hand, back when the marginal cost was a career — is catalogued in the Fires series’ history of in-game griefing and virtual scams.

The existing corpus, not just new posts, is already machine-made in parts. “A Shocking Amount of the Web is Machine Translated” (Thompson et al., AWS AI Labs and UC Santa Barbara; ACL Findings 2024) shows that multi-way parallel, machine-translated content dominates translations in lower-resource languages and makes up a large fraction of total web content in those languages, with a selection bias toward low-quality English pumped en masse through machine translation (ACL Anthology; arXiv). This is a peer-reviewed measurement, not a detector guess. The flood reaches the reference layer too: a 2024 preprint measures a rise of AI-generated content within Wikipedia (arXiv) — the adjacency that ties this to reference capture.

The number to be careful with. The widely-cited claim that “90% of online content will be synthetically generated by 2026” is not a measurement. It traces to a Europol Innovation Lab report characterizing an expert estimate, popularized via Futurism, and several analysts have argued the figure does not survive scrutiny (Futurism; Tech Business News). It is cited here only as an attributed forecast — a projection, not a fact — because it is exactly the kind of single scary number the discipline exists to catch.

The recursive-degradation mechanism the doom version fixates on — models training on model output until the quality drains out — is real, but narrower than advertised, and the corpus treats it in full elsewhere. See the model-collapse dossier for the Nature result, the accumulation counter-finding, and the “six Dead-Internet fears” test of which only two survived. The link between that dossier and this one is the spine: this page names the flood; model collapse names what the flood does to the machines drinking from it.

Right about the what, wrong about the who

The fringe theory made one true structural prediction and one false causal one.

The structural claim held. The commons is now substantially synthetic and substantially automated: bots are the majority of traffic, AI content farms number in the thousands, detector estimates put likely-AI text at 40 to 50 percent and up on some platforms, LLM botnets post indistinguishably from humans and defeat the detectors, and machine translation already dominates parts of the existing corpus. The causal claim failed. There is no single coordinating hand — no unified government-corporate psyop, no aliens, no staged collapse. That version remains unsupported.

What actually produced the “dead” internet is not a conspiracy but convergent incentives running at machine scale — the corpus’s recurring finding that you get capture-grade outcomes without a capturer. Stated structurally, and asserting no intent: platforms monetize engagement and ad impressions regardless of whether the engager is human, and their recommendation systems boosted AI slop to non-followers because it performed. Content farms run programmatic-ad arbitrage — generation cost collapsed to near zero while ad payouts stayed positive, so volume is the business model. And AI vendors benefit from demand for the generation tools, since the same models that write the slop are the product. Each of these is an economic position documented by the sources above, not an accusation.

Dead Internet Theory, then, is the folk-theory version of the capture thesis. It mistook an emergent property of incentives for a plan. That is the reframe worth keeping on record: the theorists were laughed at for the strong claim and were directionally right about the observable one, and the correct response is neither mockery nor vindication. It is to hold the line between the web is substantially bot-and-AI now, which is sourced above, and therefore someone did this to us on purpose, which is not. The evidence supports the first and refutes the need for the second.

A commons that is substantially machine-authored and machine-curated is also the substrate for the next capture surface. If the average sentence you read was written to a mean nobody chose, and the replies you get are LLM botnets, then the “public” whose thoughts you calibrate against is partly synthetic — the environmental precondition for cognition itself becoming the thing that gets captured, which is the subject of Quiet Autocomplete and, on the physiological end, the electrode and the feed.

The antidote is the same one the model-collapse work lands on, and it is embarrassingly simple and almost impossible to scale: fresh, un-averaged, first-person human signal — the rarest and cheapest thing on the internet, which is an original sentence.

Sourcing & honesty notes

Get updates on the Evil Robots series

Newsletter essays on AI escape, deception, and the humans who built them.