Just a few years ago, the Dead Internet Theory sounded like a dystopian movie plot. Born on niche forums like 4chan, it painted a picture of a ghost web\u2014a digital graveyard where most content is generated by bots and real human interactions make up only a tiny sliver of the whole. It was easy to dismiss as tech paranoia or a futurist fantasy. But what felt like fantasy yesterday is now backed by data\u2014patterns that were taking shape long before the current AI boom.<\/p>
It was the numbers that grounded this theory. The “2024 Imperva Bad Bot Report\u201d<\/mark><\/a>, which analyzed web traffic in 2023, delivered an alarming takeaway: nearly half (48,9%) of the entire internet was already dominated by bots. While that figure includes both malicious and helpful ones, the trend was unmistakable. The internet was becoming less human. It was a warning sign\u2014a quiet harbinger of a change that was about to hit with full force.<\/p> But that’s not all. An analysis<\/mark><\/a> by Graphite, covering 65,000 articles published between 2020 and 2025, found that more than half of them were written by AI language models<\/strong>. These aren’t one-offs. It’s a structural shift that’s redefining the web, long seen as a medium made mostly by humans.<\/p> Some people were sounding the alarm about this earlier. As noted by Kaitlyn Tiffany in \u201cThe Atlantic\u201d<\/a><\/mark>:<\/p> \u201cIn some sense, the people who believe the dead internet theory are right.\u201d<\/em><\/p> Set against the rise of automated content, her words now sound far less like a hot take and much more like a spot\u2011on diagnosis. And while the original theory was about government conspiracies and cover\u2011ups, its core intuition \u2014 that the internet is less and less made by humans \u2014 is starting to have a solid basis.<\/p> The generative AI explosion was the catalyst that turned a slow-moving trend into a full-blown avalanche. As powerful language and image models went mainstream, the core claims of the Dead Internet Theory stopped being a hypothesis and became a real, measurable phenomenon. The scale and speed of automated content creation have blown past anything we imagined, upsetting the fundamental balance between what’s authentic and what’s synthetic.<\/p> This problem isn\u2019t abstract anymore. We\u2019re seeing more and more evidence that it\u2019s becoming real. In its regularly updated report “AI-Generated News Tracking Center”<\/a><\/mark>, NewsGuard, which monitors the credibility of online sources, has already identified over 1,200 unreliable news sites operating in 17 languages that are entirely or largely generated by AI. These content farms flood search results with low-quality SEO articles, empty tutorials and fake news, all with one goal: to drive traffic.<\/p> A study<\/a><\/mark> published by Originality.ai found that 41.18% of long Facebook posts (over 100 words)<\/strong> were generated by AI language models.<\/p> That means almost every other long-form post an average user sees isn’t based on human thought, experience or opinion\u2014it’s generated by an algorithm.<\/p> Similar dynamics are playing out on X. Research from Harvard Misinformation Review<\/a><\/mark> shows that synthetic images and videos generated by tools like Midjourney V5 have racked up over 1,5 billion views<\/strong>. That’s a huge number, showing that people are consuming synthetic content unknowingly and on a massive scale.<\/p> But the problem doesn\u2019t end with the internet becoming less human. Something much more serious is starting. The internet is also becoming less useful as a dataset for training AI models. That\u2019s where the idea of model collapse<\/em> comes in\u2014the degradation of models that learn from content generated by other models rather than by humans.<\/p> As the researchers wrote in the journal Nature (Shumailov, I., Shumaylov, Z., Zhao, Y. et al.<\/em> AI models collapse when trained on recursively generated data. Nature<\/em> 631<\/strong>, 755\u2013759):<\/p> \u201cRecursive training on synthetic data leads to irreversible defects in the tails of the original content distribution.\u201d<\/em><\/p><\/blockquote> In other words: the more models learn from themselves, the faster they lose the ability to capture reality.<\/p> Meanwhile, researchers from the University of Texas at Austin, Texas A&M, and Purdue are observing a phenomenon they define as LLM brain rot<\/em><\/a><\/mark>. They claim:<\/p> \u201cModels exposed to viral, engagement-optimized content exhibit measurable cognitive decay.\u201d<\/em><\/p> When systems are trained on content optimized for reach\u2014simplified, condensed, and lacking depth\u2014they start thinking simplistically. It\u2019s a dynamic that can send AI into a self-degrading spiral.<\/p> If the internet really is tipping toward synthetic content, the consequences aren\u2019t just about the quality of what gets published. They also affect the economy, technology, and the future of AI models, which run on high-quality data as fuel.<\/p> First off, the importance of so-called ‘clean data’ is on the rise. As in, content from before the big generative AI boom. Organizations that have access to it gain a strategic advantage, because they can train models on material untouched by synthetic content. In practice, that means the quality gaps between companies may come not from algorithms, but from access to the right datasets.<\/p><\/a>Fewer people, more algorithms<\/h4>
<\/a>Content as toxic fuel for AI<\/h4>
<\/a>Consequences that go way beyond the internet<\/h4>