DEAD INTERNET THEORY
The claim is that most of the internet is machines talking to machines. Measured automated traffic did pass half of all requests, and that number is published and checkable. What it counts is crawlers and scrapers reaching servers, which is not the same quantity as who wrote the reply under your post.
What the theory says
The claim is that the internet died around 2016 and that most of what appears on it now is produced by machines. Under the strong version, human posting is a minority of online output and the rest is automated systems generating content for other automated systems to consume.
The idea has a specific origin. On January 5, 2021, a user posting as IlluminatiPirate published a thread on the Agora Road Macintosh Cafe forum titled "Dead Internet Theory: Most Of The Internet Is Fake" [2]. It gathered several years of similar posts from other boards into one argument. Kaitlyn Tiffany wrote the first mainstream press account for The Atlantic that August, and her piece is the reason the phrase spread beyond the forums that made it [1].
The measured automation numbers are real
The strongest evidence offered for the theory is also the part that is genuinely documented. Imperva has published an annual census of automated web traffic for more than a decade. Its 2024 report put automated traffic at 49.6 percent of all internet traffic during 2023, with the malicious subset at 32 percent [3]. The following year's edition reported automation passing half of all traffic for the first time in the series [4].
Nobody disputes these figures. They are produced by a commercial security vendor with a business reason to measure accurately, and competing measurements from other network providers land in a similar range.
What that number counts, and what it does not
Here is where the mainstream account and the theory separate, and it is a measurement question rather than a matter of interpretation.
Imperva counts HTTP requests arriving at web servers. A request is automated when it comes from a crawler, a scraper, a vulnerability scanner, a price-checking bot, an uptime monitor, a credential-stuffing script, or a search engine indexer. Every one of those makes requests without a person present, and all of them together are what pushed the figure past half.
None of that measures who wrote a comment. A single scraper hitting a site ten thousand times in an hour contributes ten thousand automated requests and zero posts. A person writing one long reply contributes one request and one post. The two quantities are not the same, they are not measured by the same instrument, and a rise in the first implies nothing directly about the second.
This is the distinction the theory's popular form does not preserve. The number is correct. The conclusion drawn from it does not follow from it.
What is documented about synthetic content specifically
The narrower claim, that a meaningful share of published material is machine generated, has its own evidence and it is weaker but real.
NewsGuard began tracking websites publishing news with little or no human oversight in May 2023, starting with 49 identified sites and expanding the count substantially in the following years [6]. Content farms producing machine-written articles at volume are documented and easy to verify by inspection.
On the platform side, the most-cited figure comes from litigation rather than research. Twitter's SEC filings stated that false or spam accounts represented fewer than 5 percent of monetizable daily active users, a number disputed during the 2022 acquisition dispute without either side producing an independent audit that settled it [5]. That the figure was contested in court, by parties with access to the underlying data, is a fair indication of how hard the quantity is to pin down from outside.
Coordinated inauthentic accounts are separately documented
Automated and semi-automated account networks operated for political effect are established by the public record. The report DiResta and colleagues prepared for the Senate Select Committee on Intelligence documented Internet Research Agency operations across multiple platforms using data sets the platforms provided [7]. Platforms now publish periodic takedown reports of their own.
This is real and it is narrower than the theory. Documented networks are attributable, finite, and were found because they were countable. That is a different claim from the internet having no humans left on it.
OPEN SOURCES
- [01] Tiffany, K. (2021). Maybe You Missed It, but the Internet Died Five Years Ago. The Atlantic, August 31, 2021. The first mainstream press account, which traces the theory to the Agora Road post.
- [02] IlluminatiPirate (2021). Dead Internet Theory: Most Of The Internet Is Fake. Agora Road Macintosh Cafe forum, January 5, 2021. The founding document.
- [03] Imperva (2024). Bad Bot Report 2024. Automated traffic measured at 49.6 percent of all internet traffic for 2023, with bad bots at 32 percent.
- [04] Imperva (2025). Bad Bot Report 2025. Automated traffic passed half of all internet traffic for the first time in the series.
- [05] Securities and Exchange Commission (2022). Twitter, Inc. Form 10-Q. The company stated that false or spam accounts represented fewer than 5 percent of monetizable daily active users, the estimate disputed during the Musk acquisition.
- [06] NewsGuard. AI Tracking Center. A running count of websites publishing news generated with little or no human oversight, first published in May 2023 with 49 sites identified.
- [07] DiResta, R. et al. (2019). The Tactics and Tropes of the Internet Research Agency. Report to the United States Senate Select Committee on Intelligence.
The alternative readings
Platforms profit from synthetic engagement and therefore do not hunt it hard
This is the most common alternative and it does not require anyone to have planned anything. Advertising rates, valuations, and creator payouts are all keyed to engagement counts. A platform that removes a large share of its own activity reports a smaller audience to advertisers.
The incentive is documented. X opened revenue sharing to creators in July 2023 with payouts tied to engagement from verified accounts [1], which puts money directly behind the production of activity. The proponent argument is that any company paying for engagement has a structural reason not to look too closely at where the engagement comes from.
Against it stands the fact that platforms do publish removals, and in volumes that undercut the idea of a total blind eye. Meta's quarterly adversarial threat reports disclose network takedowns with account counts and origins [2]. Skeptics also note that advertisers audit reach independently and have their own reason to detect inflation, which puts commercial pressure in the opposite direction.
A market in synthetic activity exists and has been prosecuted
The sale of fake followers and engagement is not a theory. The Federal Trade Commission settled with Devumi over selling fake followers and social media engagement in 2019, the first case the agency brought of that kind [3], and the New York Attorney General reached a parallel settlement covering the impersonation of real people's accounts [4].
Proponents read those cases as the visible portion of a much larger trade, on the argument that enforcement found the sellers who advertised openly and that the quiet remainder is unmeasured by definition. The counter is that this reasoning cannot be checked in either direction: an unmeasured remainder is compatible with any size at all, including a small one.
Automated opinion shaping rather than automated volume
A different reading holds that raw quantity is beside the point and that the useful question is influence. Under it, a modest number of synthetic accounts placed well inside real communities does more than a large number posting into empty space, which would make the internet functionally rather than numerically hollow.
The Senate-commissioned research on the Internet Research Agency supports the narrow form of this, since those operations were built on sustained persona work rather than volume. What it does not support is the extension, because that campaign was found, counted, and attributed, which is the opposite of undetectable.
The internet is filling with machine output regardless of intent
The gentlest version drops the conspiracy entirely. Nobody has to be steering anything for the result to hold, because generation is now cheap and publication is free, so machine-produced material accumulates on its own.
There is a technical finding in the theory's favor here. Shumailov and colleagues showed in Nature in 2024 that models trained on recursively generated data degrade, losing the tails of the distribution across successive generations [5]. If machine-written material becomes a large enough share of what is available to train on, the degradation compounds, which is a mechanism for the internet becoming less human in effect without any decision behind it.
Infrastructure has begun responding as though the volume is real. Cloudflare moved to blocking AI crawlers by default for new domains in July 2025 and introduced pay-per-crawl [6]. That a major network operator changed its defaults is evidence about crawler volume, though it says nothing about who is writing posts, which remains the quantity nobody has measured directly.
Human posting is now the minority
The strong claim, that most of what appears online is written by machines, has no measurement supporting it in either direction. No platform publishes an audited figure for the human-authored share of posts, and the outside estimates that exist are extrapolations from samples that the platforms alone could verify.
Proponents treat that absence as itself meaningful, on the reasoning that the companies best placed to publish the number have declined to. Whether an absent measurement is evidence of anything is the disagreement, and it is not one the available record can settle.
DECRYPTED SOURCES
- [01] X Corp. (2023). Ads Revenue Sharing for creators, announced February 2023 and opened in July 2023. Payouts keyed to engagement from verified accounts.
- [02] Meta (2024). Adversarial Threat Report. Quarterly disclosures of coordinated inauthentic behavior networks removed, including account counts by origin.
- [03] Federal Trade Commission v. Devumi, LLC (2019). Settlement over the sale of fake followers and social media engagement, the first FTC case of its kind.
- [04] New York Attorney General (2019). Settlement with Devumi and related entities over selling fake social media activity, including impersonation of real users.
- [05] Shumailov, I. et al. (2024). AI models collapse when trained on recursively generated data. Nature 631. The technical result behind the model collapse argument.
- [06] Cloudflare (2025). Announcement of default AI crawler blocking and pay-per-crawl for new domains, July 2025.
INFORMAL READER POLL, NOT DATA. WHICH EXPLANATION DO YOU FIND MOST CREDIBLE?