The Dead Internet Theory: Bots, Bad Data and a Genuinely Testable Claim

Sometime around 2021, on an obscure forum called Agora Road’s Macintosh Café, a user posting as “IlluminatiPirate” laid out an idea that had been drifting around image boards for a couple of years and gave it a name that would stick. The essay was titled “Dead Internet Theory: Most of the Internet is Fake”. Its claim was startling and, on first hearing, absurd: that at some point around 2016 or 2017 the internet had quietly died, that most of what now scrolls past us — the posts, the replies, the trending topics, the reviews — is generated by bots and algorithms rather than people, and that this was arranged deliberately to manipulate the population and sell to it. In its fullest form the theory tips into the supernatural-feeling and the unfalsifiable. But buried inside it is a narrower question that is not absurd at all, and that gets more answerable, and more uncomfortable, every year: how much of what you encounter online was actually made by a human being?

The strongest case, made in good faith

It is easy to laugh the theory off by quoting its silliest version, so let me do the opposite and build the strongest one, because that version deserves an honest hearing.

Begin with automated traffic, which is measurable and measured. The cybersecurity firm Imperva has published an annual “Bad Bot Report” for over a decade, drawing on traffic across its global network. Its 2024 report concluded that for the first time, automated traffic made up nearly half of all internet traffic, with “bad bots” — the malicious or deceptive kind, as opposed to legitimate crawlers like search-engine indexers — accounting for roughly a third of the total. Whatever the precise figure in a given year, the direction is not in dispute: a very large and growing share of what moves across the network is not a person. That is the empirical floor the theory stands on, and it is solid.

Now add generative text. Since late 2022, large language models have made it trivially cheap to produce fluent, human-sounding writing at industrial volume. Content farms use them to spin out articles by the thousand. “Made for advertising” websites, stuffed with machine-written filler and programmatic ads, have proliferated to the point where researchers at NewsGuard have catalogued hundreds of AI-generated “news” sites operating with little or no human oversight. On social platforms, coordinated networks of automated accounts amplify posts, inflate follower counts, and manufacture the appearance of consensus. Amazon has had to introduce limits on how many books a person can self-publish per day, after a flood of AI-generated titles; Kindle and other stores have wrestled openly with machine-written spam.

Then consider the reviews and the engagement metrics that quietly shape what you buy and read. Fake reviews are a large enough problem that the US Federal Trade Commission introduced a rule in 2024 specifically banning their sale and purchase. Follower counts, likes, and view figures have long been available for money from click farms. The “social proof” that platforms use to tell you what is popular and worth your attention is, in many corners, purchasable and therefore partly synthetic.

Stack these together and the good-faith version of the theory reads like this: an enormous and rising fraction of the text, traffic, and apparent human activity online is automated; much of it is designed to deceive, whether to sell, to manipulate opinion, or to game an algorithm; and generative AI has removed the last practical barrier — the cost of producing convincing fake human output at scale. On that reading, the claim “a lot of what you experience online is not made by people and is engineered to influence you” is not a conspiracy theory at all. It is a description of the advertising-technology economy, and it has receipts.

Where the strong case starts to fray

Having built the steelman as honestly as I can, I have to walk through where it does and does not hold, because the theory smuggles a much larger claim in alongside the defensible one, and the two need separating.

The measurable core is about proportion and automation. The larger claim is about intent and orchestration — that the deadening was done on purpose, coordinated, most often attributed vaguely to governments or a shadowy elite, to control the population. This is where the evidence thins to nothing. There is no documented central plan to replace the human internet with bots. What the record actually shows is a swarm of separate, mostly commercial actors — spammers, scammers, marketers, engagement farmers, and platforms optimising for time-on-site — each pursuing narrow incentives, whose combined output happens to be a more artificial internet. That is a very different thing from a designed operation. The distinction matters, because the difference between “the environment degraded through millions of uncoordinated selfish choices” and “the environment was deliberately killed by a central hand” is the difference between an emergent problem and a plot. Confusing the two is the same move that turns real, documented data harvesting into the mind-control legend of Cambridge Analytica — a genuine mechanism inflated into a purposeful conspiracy.

The other soft spot is the specific date. Adherents like to pin the “death” to around 2016–2017, but the human internet is demonstrably not gone. Billions of people still post, argue, share photographs of their children, and organise their lives online, and much of what any given person actually reads — the messages from friends, the group chats, the niche forums — remains overwhelmingly human. The strong version overshoots when it moves from “an increasing share is fake” to “it is all fake and the real people have left”. The first is measurable and true; the second is a mood dressed as a measurement, and it is unfalsifiable in the way that matters — because whatever human activity you point to as counter-evidence, the theory can simply reclassify as bots.

So the honest position holds two things at once, and refuses to collapse them. The proportion claim is real, quantified, and getting worse. The orchestration claim is unsupported. Between them sits a genuine, live uncertainty: nobody actually knows, with precision, what fraction of the content you personally consume is human-made, and that uncertainty is itself the disquieting part. I am not going to resolve it with a verdict, because it is not currently resolvable, and pretending otherwise would be its own small dishonesty.

The journey: a forum joke that aged into a thesis

The theory’s path is telling. It began as something between a shitpost and a genuine unease, on forums where a certain melancholy about the early internet’s disappearance was already the house mood. The people who first articulated it were, in large part, mourning. They remembered a web of small hand-built sites, weird personal pages, and forums full of recognisably human strangers, and they were watching that texture get paved over by a handful of algorithmic feeds serving optimised, homogenised content. The theory was partly a grief for a lost place, wearing the costume of a conspiracy.

What is unusual is that reality then walked toward the theory rather than away from it. Most conspiracy theories get less plausible as evidence accumulates. This one got more plausible, because the launch of ChatGPT and its successors in late 2022 supplied, almost overnight, exactly the mechanism the theory had been positing. Suddenly the cheap mass-production of convincing fake humans was not a paranoid projection; it was a product with an API and a monthly subscription. A forum notion from 2021 found itself, two years later, being cited in earnest by technology journalists and academics who would never have touched it before. The idea did not become more coherent; the world became more like it.

That trajectory is part of why the theory is worth taking seriously rather than mocking. It is a rare case where the folk anxiety anticipated a real technological shift, even if it wrapped that anticipation in cosmic language it could not support.

The platforms themselves have quietly conceded the ground the theory was pointing at. Elon Musk’s protracted 2022 attempt to escape his agreement to buy Twitter turned, in large part, on a public argument over how many of the platform’s accounts were fake — the company said under five per cent, Musk’s team claimed far more, and no outside party could settle it, which was itself the revealing part. YouTube engineers have spoken, half in jest, of an internal fear from years ago that so much of the site’s traffic was bots that the automated activity might one day be mistaken for the baseline of “real” — a scenario nicknamed, inside the company, “the Inversion”. When the operators of the largest platforms cannot themselves say with confidence how much of their traffic is human, the theory’s central question stops looking like paranoia and starts looking like a genuine measurement problem that even the insiders have not solved.

What it is really about

The dead internet theory is, underneath the bots and the timestamps, a story about loneliness and trust. Its emotional engine is the creeping suspicion that the crowd around you is not real — that the reviews you rely on, the consensus you sense, the strangers you argue with, might be phantoms wearing human masks. That is an ancient fear given a modern surface. It is the uncanny valley scaled up to an entire information environment, and it produces a specific kind of vertigo: if you cannot tell whether the voice replying to you is a person, then every online interaction carries a faint doubt that never fully resolves.

The theory answers that vertigo the way conspiracy theories usually do, by converting a diffuse, structural condition into a story with agency behind it. This is the same reframing reflex that powers grander cosmic theories like the simulation hypothesis, where the unsettling texture of modern life gets resolved by positing a designer behind the curtain. It is more bearable to believe that someone killed the internet on purpose than to accept the harder truth, which is that no one did — that the hollowing-out is the aggregate exhaust of an economy that rewards engagement over authenticity, and that there is no villain to defeat because the incentive is the villain and we are all, in small ways, inside it. A plot can be exposed and stopped. An incentive structure just grinds on.

What the theory gets genuinely right is the instinct to ask a question most people never think to ask: who, or what, actually produced the thing in front of me, and what does it want? That is a healthy reflex, and the environment increasingly rewards it. The wild version, with its central hand and its exact date of death, tells you to despair or to hunt for the puppet-master. The measured version asks you to stay curious about the provenance of what you read, to notice the tells, and to hold the honest uncertainty about how much of the digital crowd is real — an uncertainty that, for now, no one can fully dissolve, and that is likely to grow before it shrinks. The internet is not dead. But the question of how much of it is alive is a fair one to keep asking, and it deserves a better answer than either mockery or dread.