The dead internet theory 🫥

In January 2021, anonymous poster IlluminatiPirate published an essay on a small online forum called Agora Road, Dead Internet Theory: Most of the Internet Is Fake.

The theory goes that the internet died several years earlier and much of what appeared to be human activity online was generated by bots, algos, and artificial intelligence. Not the first conspiracy to be published on the internet, but it was ahead of its time, years before generative AI was producing convincing text and images at anything like today’s scale.

I’m not a robot ✅

Five years later, the theory seems less outlandish. Consider web traffic. Thales-owned cybersecurity firm Imperva estimated automated traffic exceeded human traffic for the first time in 2024. Its 2025 Bad Bot Report put bots at 51% of total internet traffic, comprising 37% ‘bad bots’ and 14% benign ones. In 2013, when Imperva began measuring the figure, bots accounted for 42.3% of traffic.

Today there are more than 6 billion humans online. It’s anyone guess as to how many bots there are. That does not mean there are now more bots than people online. One bot can generate thousands of requests. All we can measure is the activity.

The internet has been trying to distinguish people from machines for almost as long as the internet has been a thing. CAPTCHA emerged around the turn of the millennium after bots began overwhelming chat rooms and online polls. In one early case, computer programs cast around six billion votes in an online poll before it was abandoned. AltaVista later said an early CAPTCHA-style system reduced spam submissions by more than 95%.

Face/Off

Generative AI has cut the cost of producing text, photos, music, and video. So, naturally, this type of content is all over the web. Pew Research Center looked at nearly 490,000 English-language webpages published between 2021 and July 2026. In its latest sample, one in 10 pages showed significant signs of being written or edited by AI. Among pages published since ChatGPT launched in late 2022, the figure was around 35%.

One 2025 analysis created a YouTube account and found that 104 of the first 500 videos recommended to it were AI slop. The researchers identified 278 channels devoted to synthetic material which had accumulated more than 63bn views and an estimated $117m of annual revenue.

Facebook is where AI-generated nonsense has come into its own. Researchers examined 120 Facebook pages that had each posted at least 50 AI-generated images. Some were linked to spam operations. Some were scams. Collectively, the images attracted hundreds of millions of interactions. One AI-generated picture was among Facebook’s 20 most-viewed pieces of content in the third quarter of 2023, with around 40 million views.

FakeTok

Other research found a quarter of the top TikTok results examined contained synthetic imagery, with more than 80% of that material coming from accounts set up to produce AI content at scale. In a later study, 354 such accounts generated more than 43,000 mostly AI-made posts and 4.5bn views in a single month.

The problem is not only the synthetic content. Likes, comments, reviews, and followers are supposed to tell us what other people think. Once those can be gamed, the internet begins to lose one of its useful functions: social proof, a phenomenon in which people look to others to determine how to behave.

Would you like to know more?

Machines are reading more of the web too. Cloudflare (NET) found that “user action” AI crawling, where an AI system fetches a webpage in response to a user’s question, increased more than fifteenfold during 2025. AI crawlers were also the user agents most frequently blocked across entire websites via robots.txt.

Pew Research Center analysed 68,879 Google (GOOGL) searches made by 900 American adults in March 2025. When a search page contained an AI-generated overview, users clicked on a conventional search result only 8% of the time. When there was no AI overview, they clicked 15% of the time. Links contained inside the AI answer itself were clicked in just 1% of visits. Around 26% of searches displaying an AI overview ended with the user closing the browsing session, compared with 16% for traditional search results.

All the news that's fit to scrape

The Reuters Institute found in 2026 that 10% of people globally used AI chatbots for news each week, rising to 17% among the youngest respondents. Across 27 countries, just 4% of respondents said they often or always clicked through from an AI-generated news answer to the original source, compared with 19% from search and 17% from social media.

The handover to machines started well before generative AI. Yahoo’s directories were organised for people to browse. Google used software to decide which pages mattered. Social networks went further, using algorithms to decide which posts people would see.

There’s a problem if machines consume material produced by other machines. A 2024 paper demonstrated what researchers call “model collapse”. Successive generations of AI models were trained on data produced partly by earlier models. Over time, the models became less accurate, with unusual or less common information disappearing first.

Supply chain

If the web becomes filled with synthetic content, human-created information becomes more valuable. Reddit (RDDT) disclosed before its flotation that data-licensing agreements signed in 2024 had an aggregate value of $203m over two to three years. Thanks to hungry, hungry AI, human conversations are now a valuable asset class all of their own.

When supply explodes, value tends to migrate towards that which remains scarce. Authenticity could become one of those things. Communities containing verified human participants may become more attractive. Live events, expert judgement, proprietary information, and physical experiences may all command an authenticity premium.

One initiative is trying to determine where a piece of content came from, a chain of custody if you will. The C2PA, backed by firms including Adobe (ADBE), Microsoft (MSFT), Google, and Meta, is developing standards that attach verifiable information to digital media showing its source and editing history.

We’re not in Kansas anymore

Brits spent four and a half hours online a day in 2025. More than half of that time was spent on services owned by Alphabet and Meta (META). The web is now less an open plain than a series of linked, walled gardens presided over by Big Tech firms.

When Tim Berners-Lee outlined his vision for the World Wide Web at CERN in 1989, he imagined a system for people to share information, where knowledge moved freely between institutions and across borders. We’re a long way from that now. The internet has become busier while feeling emptier.

IlluminatiPirate was never identified, and the 2021 Agora Road post appears to have drawn on older discussions from imageboards and forums where authorship was often obscure. The OG internet was messy, sure, but much of it was still made by people talking to other people.