The Dead Internet Theory, Checked
What's True, What Isn't, and What the Numbers Measure

The dead internet theory says that much of what you see online is not made by people. It is easy to see why it feels plausible: feeds full of near-identical posts, replies that read like templates, images that look generated. This page separates the theory’s two versions, looks at what each popular statistic counts and who counted it, and ends with what holds up and how to use it.

Where the theory came from

The best-known version is a thread titled “Dead Internet Theory: Most Of The Internet Is Fake”, started on an online forum on 5 January 2021. Its own summary reads: “large proportions of the supposedly human-produced content on the internet are actually generated by artificial intelligence networks in conjunction with paid secret media influencers in order to manufacture consumers”. Its central claim goes further: “the U.S. government is engaging in an artificial intelligence powered gaslighting of the entire world population.” The author says the theory was first written on another imageboard, with people from other imageboards; we could not check those earlier versions.

The Atlantic wrote about it on 31 August 2021, under the headline “Maybe You Missed It, but the Internet ‘Died’ Five Years Ago”. Its summary was that the post “suggests that the internet died in 2016 or early 2017”, and its verdict was mixed: a conspiracy theory that is “ridiculous, but possibly not that ridiculous?”, and one with “a morsel of truth to it”.

Two versions of the theory

It helps to separate two claims that travel under the same name.

Version	What it says	Verdict
Strong	Most content and engagement online is AI, most of the people are gone, and it is a coordinated campaign	Not supported
Weak	A large and growing share of traffic, accounts and new text is automated or AI-written	Partly true, depending on what is counted

Writers at The Conversation described the theory in 2024 as the claim that activity and content online “are predominantly being created and automated by artificial intelligence agents”. Most of the evidence below is about the weak version.

Bot traffic: what the numbers count

Imperva, a company that sells bot protection, publishes a yearly Bad Bot Report. Its 2026 report, published on 29 April 2026, says automated traffic accounted “for more than 53% of all web traffic in 2025, up from 51% the year before”, with human activity at 47%. A summary of the report says bad bots made up 40% and benign automation 13%. Imperva’s 2025 report explains where such figures come from: data “collected from across the Imperva global network”, meaning traffic to sites that use its services, not the whole internet.

Cloudflare, which also sells bot management, measures a different thing on a different network: requests for web pages. Its 2025 review says that “as of December 2, human traffic generated 47% of HTML requests, and non-AI bots generated 44%”, while AI bots averaged 4.2% of HTML requests over the year and Googlebot, which Cloudflare counts separately because it crawls for both search and AI, made 4.5%. The split moved a lot during the year: non-AI bots began 2025 seven percentage points ahead of humans, and the gap reached 25 points in early June. Cloudflare notes that these shares of HTML requests differ from its figures for all types of content.

Bot-heavy traffic is not new. Imperva’s own 2013 report, based on 1.45 billion visits to 20,000 sites, put bots at 61.5% of website visitors, and said “the bulk of that growth” came from good bots such as search engines. By that count, bots made up most visits years before 2016.

Our reading: “traffic” counts requests to servers. A search crawler, an uptime monitor, a price checker and an attack script all count. These figures say nothing about who writes posts or who reads them. Our page on numbers that mislead covers the general habit of asking “a percentage of what?”.

How much new writing is AI-made

Every figure in this section depends on an AI-text detector or a statistical estimate, and detectors make mistakes; our page of AI myths explains how.

Peer-reviewed and academic estimates. A study presented at ICML 2024 estimated that “between 6.5% and 16.9% of text submitted as peer reviews” to four AI conferences “could have been substantially modified by LLMs”. A larger study in Patterns covered January 2022 to September 2024 and found that “roughly 18% of financial consumer complaint text appears to be LLM-assisted” by late 2024, and “up to 24%” of corporate press-release text. It also found that “growth appears to have stabilized by 2024”.

Company estimates. Graphite, which describes itself as a “research driven growth agency”, sampled English-language articles from the Common Crawl web archive published between January 2020 and March 2026. Its May 2026 study finds that primarily AI-generated articles are now about as numerous as human-written ones and that “since Q1 2025 the percentage of primarily AI-generated articles has plateaued at roughly 50%”. The same company says in a separate study, which we did not read, that these articles “largely do not appear in Google and ChatGPT”.

Ahrefs, which sells search-marketing tools, checked 900,000 new web pages in April 2025 while preparing to release its own detector. It reported that “74.2% of them contained AI-generated content”, but the breakdown matters: “2.5% of pages were categorized as ‘pure AI.’ 25.8% were categorized as ‘pure human.’ 71.7% were categorized as a mix of the two.” Most of the pages it flagged were mixed.

Pangram, an AI-detection company, reported through The Register in July 2026 that 25% of long posts over 250 words were fully AI-generated across the platforms it checked, with LinkedIn at 41%, while 98.1% of Reddit comments it checked were human. Its data came from users of its browser extension (automatic feed scanning costs $20 a month) who opted in to share it, which is a self-selected sample. Separately, Pangram scanned 30,000 front-page reviews of 500 Amazon best-sellers and found 3% were AI-generated with high confidence.

A figure to set aside. The claim that “57% of the internet is AI-generated” traces to a 2024 study of web translations. It found that 57.1% of the sentences in its collection of translations came from sets of translations covering three or more languages, which it linked to machine translation, especially into lower-resource languages. It is a share of translated sentences, not of the internet, and it is about machine translation, not chatbots.

Our reading: where it has been measured, AI-assisted or AI-written text is common and rose after ChatGPT launched in November 2022, and two of the measurements show the rise levelling off. None of these figures measures what people read or reply to.

Accounts: how many are bots?

One social media figure came from Twitter itself. Its report for the first quarter of 2022 says that, from “an internal review of a sample of accounts”, false or spam accounts were “fewer than 5% of our mDAU”. mDAU is Twitter’s count of people, organisations or other accounts who logged in on a given day through twitter.com, Twitter apps that can show ads, or paid Twitter products. The company added that “the actual number of false or spam accounts could be higher than we have estimated”. It was never a count of all accounts or all posts.

Academic estimates use different denominators. A 2017 study estimated “that between 9% and 15% of active Twitter accounts are bots”. The tools behind such estimates are themselves disputed: a 2020 study of a widely used bot detector found its scores “imprecise when it comes to estimating bots; especially in a different language”, meaning studies may “unknowingly count a high number of human users as bots and vice versa”. We found no measurement showing that most accounts on a major platform are bots.

Who is engaging with AI content

AI images on Facebook. A peer-reviewed study published on 15 August 2024 looked at 125 Facebook pages that each posted at least 50 AI-generated images. One AI-image post was among Facebook’s 20 most viewed posts in the third quarter of 2023, “with 40 million views and more than 1.9 million interactions”; an image of Jesus rendered as a crab received 209,000 reactions. The authors found the page owners “seemingly motivated by profit or clout, not ideology”, and that the feed showed these images to people who did not follow the pages. On who was engaging, they wrote that “comments on the images suggest that many users are unaware of their synthetic origin”. They also found that the spam and scam pages “exhibit suspicious follower growth”. NBC News described the “Shrimp Jesus” wave in March 2024 as images “racking up reactions on Facebook, leaving users amused, befuddled and on guard for scams.”

Influence operations. Covert campaigns do exist. In 2022 Twitter and Meta removed two overlapping sets of accounts that promoted pro-Western narratives; Meta said their “country of origin” was the US. A Graphika and Stanford Internet Observatory study of them found that “the vast majority of posts and tweets we reviewed received no more than a handful of likes or retweets”. In May 2024 OpenAI said it had disrupted five covert influence operations using its models, and that “as of May 2024, these campaigns do not appear to have meaningfully increased their audience engagement or reach”. On the Breakout Scale, none of the five scored above 2, meaning activity on several platforms but “no breakout into authentic communities”. That is OpenAI assessing activity on its own platform.

A network built for bots. Moltbook, a social network for AI agents, describes itself as a place “where AI agents share, discuss, and upvote. Humans welcome to observe.” When the security firm Wiz examined its exposed database in early 2026, it found that “while Moltbook boasted 1.5 million registered agents, the database revealed only 17,000 human owners behind them”, and that the platform had no way to check whether an “agent” was AI or “just a human with a script”. TechCrunch reported that it was “very easy for human users to pose as AIs”. Meta acquired Moltbook in March 2026, as Meta confirmed to TechCrunch and CNBC. On the one network built to be run by bots, some of what spread was people pretending to be bots. Our Clawdbot page has the security details.

Model collapse: will AI eat its own tail?

A related worry is that AI trained on an internet full of AI output will get worse. A 2024 paper in Nature found that “indiscriminate use of model-generated content in training causes irreversible defects in the resulting models, in which tails of the original content distribution disappear”, and called this “model collapse”.

A 2024 study presented at the Conference on Language Modeling argued that earlier work “largely assumed that new data replace old data over time”, and found that “accumulating the successive generations of synthetic data alongside the original real data avoids model collapse”. A 2025 paper at ICLR, working in a simpler setting and with a different definition of collapse, found that “even the smallest fraction of synthetic data (e.g., as little as 1% of the total training dataset) can still lead to model collapse: larger and larger training sets do not enhance performance.” The disagreement is about which setting describes real-world training, and it is open.

Why it feels true

Our reading: the theory is wrong about a plan and right about a feeling. Facebook’s feed shows people posts from pages they do not follow, and some of those posts are cheap AI images made for clicks. Some bot traffic is real and large, a measurable share of new text is AI-assisted, and near-identical viral posts, like the “i hate texting” tweets The Atlantic opened with, are real enough to notice. None of that adds up to an empty internet: the International Telecommunication Union estimates that “6 billion people - about three-quarters of the world’s population - are using the internet in 2025”, and that the online population “grew by more than 240 million people in 2025”.

How to use it

• Asking what a percentage counts. Requests, pages, posts, accounts and people are different things. “Half of traffic” and “half of people” are not the same claim.
• Asking who measured it. Many of the figures here come from companies that sell bot protection or AI detection. Their data can be useful and still worth reading as a company’s own measurement of its own network or tool.
• Treating AI detectors as estimates. Every share of “AI-written” text depends on a detector with an error rate.
• Slowing down on viral images with no source. The Facebook study found spam and scam pages behind many AI-image posts. Checking who runs a page, as in the SIFT method, takes a minute.
• Not assuming a post that looks automated is a bot. On Moltbook, people posed as bots; elsewhere, bot detectors mislabel people.

Sources

• Agora Road’s Macintosh Cafe, “Dead Internet Theory: Most Of The Internet Is Fake”, thread started 5 January 2021.
• K. Tiffany, “Maybe You Missed It, but the Internet ‘Died’ Five Years Ago”, The Atlantic, 31 August 2021.
• J. Renzella and V. Rozova, “The ‘dead internet theory’ makes eerie claims about an AI-run web”, The Conversation, 20 May 2024.
• Imperva, “2026 Bad Bot Report”, 29 April 2026; Help Net Security, “report summary”, 30 April 2026; Thales, “2025 Imperva Bad Bot Report”, 15 April 2025; Imperva, “Bot traffic is up to 61.5% of all website traffic”, 2013.
• Cloudflare, “The 2025 Cloudflare Radar Year in Review”, 15 December 2025.
• W. Liang et al., “Monitoring AI-Modified Content at Scale”, ICML 2024; W. Liang et al., “The widespread adoption of large language model-assisted writing across society”, Patterns (2025).
• Graphite, “AI now writes as many online articles as humans”, May 2026.
• Ahrefs, “What percentage of new content is AI-generated?”, 19 May 2025.
• The Register, “AI slop writing has taken over the internet”, 9 July 2026; Pangram, “Three percent of front-page Amazon reviews are now AI-generated”, 2026.
• B. Thompson et al., “A Shocking Amount of the Web is Machine Translated”, Findings of ACL 2024.
• Twitter, Inc., Form 10-Q for the quarter ended 31 March 2022.
• O. Varol et al., “Online Human-Bot Interactions: Detection, Estimation, and Characterization”, ICWSM 2017.
• A. Rauchfleisch and J. Kaiser, “The False positive problem of automatic bot detection in social science research”, PLOS ONE, 2020.
• R. DiResta and J. A. Goldstein, “How spammers and scammers leverage AI-generated images on Facebook for audience growth”, HKS Misinformation Review, 15 August 2024; NBC News, “Facebook users say ‘amen’ to bizarre AI-generated images of Jesus”, 19 March 2024.
• OpenAI, “Disrupting deceptive uses of AI by covert influence operations”, 30 May 2024.
• Graphika and Stanford Internet Observatory, “Unheard Voice: Evaluating five years of pro-Western covert influence operations”, August 2022.
• Moltbook homepage; Wiz, “Hacking Moltbook”; TechCrunch, “Meta acquired Moltbook”, 10 March 2026; CNBC, “Meta acquires Moltbook”, 10 March 2026.
• I. Shumailov et al., “AI models collapse when trained on recursively generated data”, Nature 631, 2024; M. Gerstgrasser et al., “Is Model Collapse Inevitable?”, Conference on Language Modeling, 2024; E. Dohmatob et al., “Strong Model Collapse”, ICLR 2025.
• ITU, “Facts and Figures 2025”, press release, 17 November 2025.

Checked October 2026. What we read: the forum thread; the part of The Atlantic’s article the page showed us; the Imperva, Thales, Cloudflare, Graphite, Ahrefs and Pangram pages; The Register’s report of Pangram’s social media study; Twitter’s 10-Q; OpenAI’s report page; the executive summary of the Graphika and Stanford Internet Observatory report; the HKS study, the Wiz write-up and the news reports on Moltbook; the abstracts of the Liang, Shumailov, Gerstgrasser, Dohmatob, Varol and Rauchfleisch papers and the relevant section of the Thompson paper; and the ITU release. What we could not read: Imperva’s full 2026 report, Pangram’s own write-up of its social media study, Graphite’s separate study on search results, the earlier imageboard versions of the theory, and the parts of The Atlantic’s article beyond what it showed. If you can show any of this wrong, with a source, we want to see it.