The Dead Internet Theory: Is the Internet Becoming Less Human?

The internet has never had more people on it. It has also never had more machines. That sounds like a contradiction, but it isn’t. Billions of people use the internet every day while automated systems crawl websites, index pages, scrape databases, place advertisements, detect fraud, recommend content and increasingly generate text, images and video. AI systems are now joining that machinery. They can read websites, summarize them, generate new material and, in the case of increasingly capable agents, interact with websites on behalf of a user.

This is where the Dead Internet Theory comes in. The theory, in its strongest form, argues that much of the internet is no longer populated by genuine human activity. Instead, bots and AI agents supposedly generate a large proportion of what people see, while governments and corporations manipulate the resulting environment to manufacture consensus, engagement and public opinion. There is no reliable evidence for that version of the theory.

There is, however, a much more interesting question underneath it. The internet can remain overwhelmingly populated by real people while becoming increasingly dependent on machines to create, rank, distribute, summarize and consume information. Some of the measurements coming out of major internet infrastructure providers already show automated requests exceeding human requests. Academic research has found that AI-generated or AI-assisted text now appears in a substantial share of newly published websites. Search engines are increasingly answering questions without sending users to the original sources, and AI crawlers are consuming websites for training and search at enormous scale. The internet is not dead. But the way humans encounter it is changing.

The theory existed before Generative AI

Dead Internet Theory did not begin with generative AI. The idea developed through anonymous online communities before being popularized in its recognizable modern form by a 2021 post on Agora Road’s Macintosh Cafe. The post argued that the internet had become increasingly artificial, with automated activity replacing genuine human interaction and powerful institutions supposedly benefiting from that transformation. The theory later spread beyond obscure internet communities and became a recurring subject of discussion in mainstream technology and culture coverage.

In 2021, ChatGPT did not yet exist as a public product. Large language models were already capable of generating text, but they had not become a routine part of everyday publishing. The web was nevertheless already full of automated activity. That fact matters because it separates two developments that are often mixed together. The internet did not become automated because of generative AI. It was already heavily automated. Generative AI has made that automation considerably more capable.

Imperva’s 2016 research, based on more than 16.7 billion visits across 100,000 randomly selected domains, estimated that bots accounted for 51.8% of the traffic in its dataset. Human traffic represented 48.2%, while 22.9% came from what Imperva classified as good bots and 28.9% from bad bots.

That did not mean that more than half of internet users were robots. The measurement concerned traffic, not people. A crawler can generate thousands of HTTP requests while representing no human user at all. It did demonstrate something that is easy to overlook when discussing the Dead Internet Theory: machines were already responsible for an enormous amount of internet activity long before the AI boom.

The number that sounds like proof of Dead Internet Theory

Imperva’s more recent figures are even more striking. Its 2025 report found that automated traffic accounted for 51% of web traffic in 2024, meaning automated requests had overtaken human traffic in the company’s measurement for the first time in a decade. The 2026 report, based on 2025 data, puts automated traffic above 53%, with human traffic at 47%. Imperva also reports that 37% of total internet traffic in 2025 came from bad bots.

Cloudflare reported another milestone in July 2026. Its network telemetry showed automated systems generating roughly 57% of web requests, compared with about 43% attributed to humans. Cloudflare described this as the first time automated bot traffic had eclipsed human activity in its measurement.

If those numbers are presented without context, they sound like the Dead Internet Theory has been proven. They haven’t. The important word is requests. A web request is an interaction with a server. It is not a person. A human opening a webpage might cause dozens of requests for HTML, JavaScript, CSS, images, fonts and other resources. A crawler can make requests continuously and systematically across thousands of pages. An AI training system can retrieve enormous amounts of information without a human sitting behind each request.

There is currently no global measurement capable of telling us that, for example, 57% of the people using the internet are bots. Cloudflare is measuring requests passing through its network. Imperva is measuring traffic in its own dataset. Their numbers tell us that machine-generated web activity has become enormous. They do not tell us that humans have become a minority of internet users.

That is why the claim “bots now make up most of the internet” needs to be treated carefully. A more accurate statement is that automated systems now generate a majority of web requests in some large-scale infrastructure measurements. That is still an extraordinary development. It simply isn’t the same thing as saying that most people online are fake.

Bots have changed from crawlers into participants

For most of the web’s history, automation was relatively narrow. Search engines crawled pages. Monitoring systems checked whether websites were available. Scrapers collected information. Security tools scanned applications. Spam networks created accounts and distributed unwanted material. These systems did not need to behave like people. The new generation of AI-enabled systems is different.

An AI agent can interpret a page, extract information, decide what to do next and interact with another system. It can search for a product, compare options, fill in a form, retrieve information from several sources and potentially complete an action. That makes the distinction between a “bot” and an “internet user” increasingly complicated.

Imperva describes AI agents as a new class of internet participant because they can retrieve information, execute workflows and act on behalf of users. The company’s 2026 report says 27% of bot attacks targeted APIs and identifies financial services as the sector accounting for 24% of bot attacks and 46% of account takeover incidents in its 2025 dataset.

This is an important cybersecurity development, but it also has consequences for the broader internet. A conventional crawler asks a website for information. An agent can ask a website to do something. That difference becomes particularly important when thousands or millions of agents begin interacting with websites at machine speed.

Cloudflare now expects humans to become a rounding error

The numbers being measured today are already striking, but Cloudflare executives are now making a much more aggressive prediction about what comes next.

In an August 7, 2026 report, The Register quoted a Cloudflare executive saying that humans could eventually become a “rounding error” on the internet as machine-generated traffic expands. The company expects machine-generated traffic to increase by roughly 1,000 times over the next five years. The prediction was made against the backdrop of rapidly growing AI-agent activity, rather than a claim that the number of human internet users is about to collapse.

Cloudflare has already reported that automated systems account for more than half of web requests in some of its current measurements. The new prediction concerns what happens when today’s relatively simple automated requests are joined by large numbers of AI agents that can independently browse, search, retrieve information and perform tasks. The difference could be enormous.

A human browsing a website generally moves through pages at human speed. An agent performing a research task can retrieve many pages in seconds. An agent comparing products can query multiple websites, follow links and repeat the process without waiting for a person to read each page. A fleet of agents performing these tasks simultaneously can therefore generate vastly more requests than the humans who initiated them.

This creates a mathematical problem with the way internet traffic is commonly discussed. If one person initiates a task that produces hundreds of machine requests, the number of requests can become overwhelmingly non-human even though the underlying demand still comes from a human.

Cloudflare’s forecast should therefore not be interpreted as “humans will disappear from the internet.” It means that the volume of machine-to-machine communication could become so much larger than direct human-generated traffic that human requests become comparatively insignificant as a share of total network activity.

Today, a human might search Google, open several websites and decide what to buy. Tomorrow, the human may simply tell an AI agent what they want. The agent could search the web, compare products, check availability, read reviews, negotiate through an API and complete the transaction. The human remains the decision-maker. The network traffic no longer looks human.

This is also why Cloudflare’s prediction should be treated as a scenario rather than a settled fact. A thousand-fold increase over five years would require sustained growth in agent adoption, infrastructure capacity and machine-to-machine applications. It would also depend on how Cloudflare defines and measures machine traffic as the composition of the web changes.

The prediction is nevertheless useful because it exposes where the current trend could lead. If AI agents become a routine interface between people and the web, the dominant consumer of many websites may no longer be the person sitting in front of the screen. It may be software acting on that person’s behalf.

That would represent a much deeper change than simply having more bots. The web was designed around an implicit assumption that a human was usually on the other side of a request. An agentic internet weakens that assumption. A website may increasingly need to distinguish between a person reading a page, a search crawler indexing it, an AI model training on it and an autonomous agent acting on behalf of a customer.

In that environment, “human traffic” becomes only one category of internet activity rather than the default and that brings the Dead Internet Theory closer to an uncomfortable technical reality. The question may eventually stop being whether bots outnumber humans. It may become whether the majority of meaningful activity on the web happens directly between people at all.

The web is also being read by machines that want to create more information

The next change is more difficult to see. Search engines have always crawled the web because they need an index. AI companies increasingly crawl it because models need data and AI-powered systems need information.

Cloudflare has documented significant growth in AI and search crawler activity. Its analysis found AI and search crawler traffic increased 18% between May 2024 and May 2025 after accounting for customer growth, and 48% when new customers were included. The important difference is what happens after the crawl. A conventional search crawler indexes a page so that a user can later find the original source. An AI system may retrieve information from that page and incorporate it into a generated answer. The user may never visit the original website.

That creates a new relationship between the publisher, the crawler and the reader. The website can still be extremely valuable to an AI system while receiving less direct traffic from human readers. This is one of the central tensions of the emerging AI web. The machines may be consuming more of the internet precisely while humans are consuming less of the original material.

AI-generated content is no longer a marginal phenomenon

There is now evidence that generative AI is changing the composition of newly published web content. A 2026 study by Jonas Dolezal, Sawood Alam, Mark Graham and Maty Bohacek used Internet Archive data to construct a representative sample of websites published between 2022 and 2025. The researchers applied an AI text detector to estimate the prevalence of AI-generated and AI-assisted text.

By mid-2025, approximately 35% of newly published websites in their sample were classified as containing AI-generated or AI-assisted text. Before the launch of ChatGPT in late 2022, the measured proportion was effectively zero under their methodology.

That is one of the strongest pieces of evidence relevant to the modern interpretation of Dead Internet Theory. It also needs to be described precisely. The study did not conclude that 35% of every webpage on the internet is AI-generated. It examined newly published websites. It also grouped AI-generated and AI-assisted material together.

There is an important difference between a journalist using an AI system to clean up grammar and an automated content farm publishing thousands of articles without meaningful human review. The study’s methodology cannot turn that distinction into a single global number. Nevertheless, the direction is difficult to ignore. AI is no longer merely helping a small number of writers experiment with text. It has become part of the production process for a substantial amount of new web content.

The more interesting finding is not accuracy. It is similarity.

The same 2026 study found that increasing levels of AI-generated text were associated with lower semantic diversity and higher positive sentiment. It did not find statistically significant evidence that increasing AI-generated text reduced factual accuracy or stylistic diversity.

A common description of the AI-content problem is that the internet is going to fill up with incorrect information. That may happen in particular cases, but the available research does not justify saying that AI-generated content has already made the web broadly less accurate.

A different problem is emerging. Large language models learn statistical patterns from enormous quantities of existing text. When people use the same models to produce new material, those statistical preferences can appear repeatedly across otherwise unrelated websites.

The result can be an internet that contains a large number of different pages while becoming less diverse in the information those pages express. That is a quieter form of homogenization. It is also one of the reasons that some AI-generated writing can feel strangely familiar even when the subject is new.

Social media is not mostly bots

The evidence becomes considerably weaker when the claim is extended from web traffic to social-media users. There are enormous numbers of fake accounts and automated systems on social networks. Platforms routinely remove coordinated inauthentic behavior, spam networks and automated accounts. Researchers have repeatedly demonstrated that bots can influence the apparent popularity of particular topics.

That does not establish that most social-media users are artificial. The distinction between traffic and participation matters again. A crawler can generate millions of requests without having a social-media account. A single automated account can generate thousands of posts. A human can operate several accounts. A human can also use AI to write a post. The categories overlap.

This is one reason claims such as “most social-media posts are written by bots” require platform-specific evidence rather than extrapolation from web-traffic statistics. Research on Reddit provides a useful example. A study examining more than two years of activity across 51 subreddits found that machine-generated text was generally a relatively small component of total activity, although some communities reached levels of up to about 9% during particular periods.

That is substantial enough to matter. It is nowhere near evidence that Reddit has become a machine-only conversation. The same lesson applies more broadly. A small number of highly productive automated accounts can produce a disproportionate amount of visible material. The percentage of automated accounts and the percentage of visible activity they generate are not necessarily the same.

Popularity can be manufactured without manufacturing the people

This is where recommendation algorithms enter the story. Suppose every post on a platform is written by a real human. The platform still has to decide which posts users see. That decision is generally automated. Recommendation systems rank content using signals that can include predicted engagement, viewing behavior, interaction history, relevance and other platform-specific factors. A person therefore does not experience the complete set of things that other humans posted. They experience a machine-selected subset. That creates a different kind of artificiality. The content can be entirely human. The distribution of attention is not.

This information matters because online popularity is often interpreted as evidence of public opinion. A post with ten million views can feel like a reflection of what millions of people believe. In reality, it may be the product of a ranking system optimized around engagement and retention rather than representative sampling. Bots can then exploit that system. Fake accounts can generate early interactions. Engagement farms can create artificial signals. Coordinated groups can repeatedly push particular material. Recommendation systems can react to those signals and distribute the content further. The resulting popularity may be real in the sense that millions of people actually saw the post. The initial signal that caused the distribution may not have been organic. That is a genuine problem, and it does not require the internet to be mostly bots.

Governments really do operate artificial influence networks

There is also no need to speculate about whether governments or politically motivated organizations have used fake online identities. They have.

Meta has repeatedly documented coordinated inauthentic behavior networks involving fake accounts, pages, groups and personas. Google has published regular reports on state-linked influence operations. OpenAI has documented cases where its models were used by covert influence networks.

OpenAI has reported operations using AI to generate social-media comments, articles, biographies, names, translations, research and code. One Russian-linked operation used AI-generated political comments in Russian and English. Other operations used models to create or support content intended for platforms including Facebook, Instagram and X.

A campaign operator can decide what message to promote. AI can help generate variations. Automation can create and manage accounts. Human or automated systems can distribute the material. Recommendation systems can provide additional reach. The resulting content may eventually reach a genuine human audience. That is how synthetic influence works in practice. It does not require the entire audience to be artificial.

Advertising has its own machine population

Some of the largest automated systems on the internet have nothing to do with social media or politics. They exist because advertising is automated. Modern advertising exchanges process enormous volumes of machine-generated requests. Advertisers bid on opportunities to display advertisements, systems evaluate those bids, publishers serve the ads and analytics platforms measure what happened. Fraudsters can attack almost every part of that process. Bots can generate impressions. Automated systems can create clicks. Fraud networks can simulate conversions. Malicious applications can generate advertising activity in the background.

HUMAN Security has documented operations capable of producing hundreds of millions of fraudulent advertising requests per day. Its research into the IconAds operation found as many as 1.2 billion bid requests per day at peak. Its investigation of SlopAds identified 224 applications associated with a campaign that generated more than 38 million downloads. These figures describe specific fraud operations, not the overall advertising ecosystem, but they demonstrate the scale at which machine-generated economic activity can operate. This is another reason “bot” is too broad a category. A legitimate search crawler, an AI training crawler and an advertising-fraud network are all machines. Their economic and social effects are completely different.

Search is changing the relationship between websites and readers

The most important transformation may be happening in search. For two decades, the basic web-search bargain was simple. Search engines crawled websites and sent users to them. Publishers received traffic. The search engine received a useful index and an advertising business. AI search can insert another step. A model retrieves information from several websites, synthesizes it and provides an answer directly on the search page.

Pew Research Center examined 68,879 Google searches from a panel of 900 U.S. adults and found that AI summaries appeared on 18% of the searches in its study. When an AI summary appeared, users clicked a traditional search-result link in only 8% of visits. When there was no AI summary, the corresponding figure was 15%. Only 1% of visits involved a click on a link directly within the AI summary. Users also ended their browsing session more often after seeing an AI summary, 26% compared with 16% when no summary appeared.

This does not prove that AI search is destroying the open web. It does establish a measurable change in user behavior. The original website may still be the source of the information. The AI system can extract that information, synthesize it and satisfy the user without the user ever visiting the source. That creates a difficult economic question for publishers. If machines consume a growing amount of their content while humans visit fewer of the original pages, who pays for the creation of the next generation of information? That question is a structural problem for the web.

The feedback loop between AI and the web

There is another reason the growing amount of synthetic content matters. AI systems are trained on enormous datasets. The public internet has historically been an important source of human-created text, images, code and other information. As AI-generated material becomes more common online, it becomes increasingly difficult to guarantee that future datasets contain only original human-generated material. This creates the possibility of a feedback loop. A human writes something. An AI reads it. The AI produces a summary. The summary is published. Another system collects it. A later model trains on the generated material. The next model produces another version. The process can continue.

Researchers have demonstrated that recursive training on synthetic data can produce model collapse. In a 2024 paper, researchers showed that when models were repeatedly trained on data generated by previous generations of models, information from the original distribution could progressively disappear. The effect was especially pronounced for less common or “tail” information. That result is often simplified into “AI training on AI content will make AI collapse.” The actual research is more nuanced.

A separate 2024 study found that replacing real training data with successive generations of synthetic data can indeed lead toward collapse, but accumulating synthetic data alongside the original real data avoided collapse in the researchers’ experiments. Other work has shown that verification of synthetic data can also reduce the problem. So, model collapse is a genuine technical risk. It is not an inevitable consequence of every use of synthetic data. The practical issue is whether future AI developers can maintain enough high-quality, independently sourced human data and distinguish it from synthetic material.

What “AI slop” really changes

The popular term “AI slop” usually refers to low-effort machine-generated content. The deeper concern is information quality rather than aesthetics. The value of having ten independent websites discussing an event is that they may contain ten independent observations. If ten websites publish variations of the same machine-generated explanation, the web has gained ten URLs but not necessarily ten independent sources of information.

A search engine sees pages. A model sees training examples. Neither automatically knows whether those pages represent independent observations or thousands of copies of the same generated idea.

The 2026 research finding that greater AI-generated text correlates with lower semantic diversity is therefore worth watching. It does not establish that the internet has entered a model-collapse loop, but it provides empirical evidence for one of the mechanisms that could make a synthetic information environment less diverse.

The internet is still overwhelmingly human

For all of these developments, there is a basic fact that the strongest Dead Internet claims cannot overcome. Humans are still everywhere. Billions of people use the internet. People continue to publish original reporting, conduct scientific research, maintain open-source projects, run communities, create art, upload videos and communicate directly with one another.

There is no credible evidence that most internet users are bots. There is no credible evidence that human participation has been secretly replaced. Also, there is no evidence that the majority of online conversations are actually machine-to-machine conversations disguised as human discussion. What has changed is the environment around those humans. A person can write an article. A crawler can copy it. A search engine can index it. An AI model can summarize it. A recommendation system can decide who sees that summary. Another AI system can use the information in a generated answer. An advertising system can place a bid around the resulting traffic. A fraud system can attempt to imitate the resulting engagement. All of those things can happen around a single piece of human-created information. The person remains real.

The surrounding internet becomes increasingly automated.

The real dividing line is no longer human versus machine

The original Dead Internet Theory assumes a binary. There are humans, and there are fake machines pretending to be humans. The modern internet is considerably messier. A journalist can use an AI assistant to research an article. A programmer can ask an AI system to generate code. A business can use an AI agent to answer customer questions. A human can write a Reddit post and use AI to edit it. A fully autonomous system can publish a webpage without human review.

Where exactly does “human” end and “machine” begin? That question is becoming harder to answer because AI is moving from a content generator into a general-purpose interface between people and information. It is whether humans remain responsible for its creation, verification and purpose. That is also why provenance is becoming more important. As synthetic material grows, knowing where information came from may eventually matter as much as the information itself.

So, is the Dead Internet Theory coming true?

If the claim is that humans have been replaced by bots, no. If the claim is that governments secretly control the entire internet through artificial accounts, there is no evidence for it. If the claim is that most social-media users are fake, current evidence does not support it. If the claim is that AI already generates most of the internet’s content, that has not been demonstrated. But if the theory is interpreted as an early warning that the internet would become increasingly automated, algorithmically curated and populated by synthetic content, parts of that prediction are becoming measurable.

Imperva says automated traffic exceeded 53% of its measured web traffic in 2025. Cloudflare reported roughly 57% automated web requests in July 2026. A 2026 academic study found approximately 35% of newly published websites in its sample contained AI-generated or AI-assisted text by mid-2025. Pew found that Google users were less likely to click traditional search results when an AI summary appeared. Researchers have demonstrated that uncontrolled recursive use of synthetic training data can produce model collapse, while other research shows that retaining original data and carefully verifying synthetic data can mitigate the problem. And documented influence operations have already used AI to produce and distribute political material.

None of those findings proves the theory. Together, they describe a different phenomenon. The internet is becoming increasingly machine mediated. Machines do not have to replace humans for the character of the internet to change. They only need to become responsible for a growing share of the processes through which humans encounter information. A person may still choose what to believe. But increasingly, a machine decides what that person is likely to see first. A human may still write the original article. A machine may decide whether anyone encounters it. A website may still contain information written by a person. An AI system may answer the question without sending anyone there. And a growing amount of new material may itself have been produced with the help of another machine.

That is not a dead internet. It is something more complicated: a human internet operating inside an increasingly automated information system. The question worth watching now is not whether humans have disappeared. It is how much of the internet a human actually experiences without a machine standing between the person and the information. That number is much harder to measure than bot traffic. It may also be the statistic that eventually matters most.