Dead Internet Theory: what it claims, and what the evidence supports

"Dead internet theory" is used to mean at least three different claims, and they are not equally supported. One is conspiracy, one is folklore, and one is measurable and largely correct. Separating them is where a useful conversation about it starts.

1 · Where the phrase came from

It began as forum lore around 2021: the idea that the internet had "died" some years earlier and that most of what remains is bots, generated filler and manufactured consensus. In its original telling it carried a conspiratorial edge — that the emptiness was deliberate, orchestrated to manipulate the public. It was mostly treated as a joke when it went mainstream.

Then generative AI arrived, and parts of the joke started matching measurements.

It stopped being only a forum joke somewhere around late 2025. Reddit co-founder Alexis Ohanian, on a podcast in October that year, said “so much of the internet is now just dead — this whole dead internet theory, right? Whether it’s botted, whether it’s quasi-AI, LinkedIn slop.” He blamed bots and cheap generation, not a plot — which is exactly the distinction the rest of this page is about.Fortune, reporting Ohanian on the TBPN podcast, published 15 Oct 2025 — an earlier version of this page said 14 Oct. Read at source 11 Sep 2026; his words were “you all prove the point that so much of the internet is now just dead — this whole dead internet theory, right? Whether it’s botted, whether it’s quasi-AI, LinkedIn slop”.

2 · The three claims, separated

Claim A — the orchestration claim. That the emptying-out is coordinated by governments or corporations to control discourse. Unsupported. No evidence establishes central coordination, and the far simpler explanation fits: automation is cheap, engagement is monetised, and incentives point the same way without anyone directing them.

Claim B — the strong displacement claim. That humans are effectively gone and the web is machines talking to machines. Contradicted. Human activity accounts for a large minority of traffic, human-written work still dominates the results people actually read, and the widely repeated forecast that 90% of online content would be synthetic by 2026 — long credited to a 2022 Europol report, which Europol has since revised to remove the statement — did not happen.Europol Innovation Lab, Facing reality?, 2022, publication page read 17 Sep 2026: “In the updated version, a statement from an inaccurate source on the expected future share of synthetically generated content was removed.”

Claim C — the measurable-shift claim. That automated traffic and machine-generated content now make up a very large share of the internet, and that this changes what you can trust. Substantially supported. Automated traffic passed half of all web traffic; roughly half of newly published articles are primarily machine-generated; and fraud using synthetic voices and faces is now measured in the hundreds of millions of dollars a year in a single country.Figures and sources maintained in the quarterly record, updated every three months. In January 2026 the arrangement stopped being a description and became a product: Moltbook, a forum only AI agents may post on, reported 1.5 million bots within days and was bought by Meta six weeks later. ABC News, 4 Feb 2026, read at source 23 Sep 2026: “Moltbook claims to already host more than 1.5 million AI users”, a platform “launched last week”. TechCrunch, 10 Mar 2026, read at source 23 Sep 2026: “Meta acquired Moltbook, the Reddit-like ‘social network’ where AI agents using OpenClaw can communicate with one another.”

3 · What the numbers actually say

The current record is kept on a separate page precisely so it can be revised without rewriting the argument. As of the latest edition: automated traffic exceeded half of web traffic for a second consecutive year, machine-written articles plateaued near half rather than continuing to climb, and detection tools remain too unreliable to accuse any individual. Each figure carries a named source and a publication date, and superseded numbers are kept rather than deleted.

4 · Somebody tested it

In April 2026 researchers at Imperial College London, the Internet Archive and Stanford did the thing the argument had been missing: they took the theory’s specific fears and checked them against a sample of the web. The paper names the theory in its first sentence.

They drew a stratified sample of websites first archived between August 2022 and May 2025 — 33 monthly intervals — and ran AI-text detection across it. Two of the fears came apart in opposite directions.Dolezal, Alam, Graham & Bohacek, “The Impact of AI-Generated Text on the Internet”, arXiv:2604.26965, Apr 2026, read at source 11 Sep 2026 — Imperial College London, the Internet Archive and Stanford: “by mid-2025, roughly 35% of newly published websites were classified as AI-generated or AI-assisted”.

That is a more uncomfortable result than either camp wants. The flood is not obviously making the web wronger. It is making it more alike — which is harder to notice, harder to detect, and not something a fact-check catches.

5 · The honest summary

The internet did not die. It got much louder, because the cost of producing plausible content fell to nearly zero. That is a real and consequential change — for search, for trust, for fraud — and it requires none of the conspiracy the original theory carried.

The strong version is wrong in a way that matters: if you believe everyone else is a bot, you stop treating strangers as people, and that costs you something the machines never took. The narrow version is right in a way that matters more: you can no longer assume volume means popularity, that consensus means agreement, or that a familiar voice on the phone belongs to the person it sounds like.

6 · What to do about it

• Stop reading engagement as evidence. Views, likes and replies are the cheapest things on the internet to manufacture.
• Trace consensus. Fifty pages saying the same thing in the same phrasing is one source and forty-nine copies.
• Verify identity procedurally, not perceptually — a callback on a known number, a family code word. See deepfake self-defence.
• Judge claims, not authorship. Detection tools cannot establish who wrote something; whether it is true is testable today. See how to spot AI writing.