From Conspiracy to Mainstream: Is the Internet Dying Before Our Eyes?
A decade ago, the Dead Internet Theory was a fringe conspiracy theory confined to the darker corners of the web. Now traffic figures, content-production statistics and AI research all point in the same direction. Bots are producing an ever-larger share of the web’s traffic, while human-written text is gradually becoming a scarce resource. The question is no longer whether the internet is changing, but what will be left of it.
Text by Martti Asikainen, 6.8.2026 | Photo by Adobe Stock Photos
Mark Twain, one of the sharpest pens of his age, famously told a New York Journal reporter in 1897 that the reports of his death were greatly exaggerated. The same can no longer be said of the internet, which we can see fading before our eyes. This came to mind recently when I read an enthusiastic post by a Finnish business executive describing how he had automated an AI to open Reddit for him four times a day and respond to eight discussions relevant to his industry. This is no longer the exception; it is becoming a common way of maintaining a presence online.
The Dead Internet Theory has been circulating on online forums since the 2010s. At first, the idea that the internet was now dominated by non-human activity was dismissed as a conspiracy theory. Since then, it has become a measurable fact. Several companies specialising in cybersecurity and web traffic agree not only that bots now outnumber humans online, but also that pinpointing exactly when this tipping point occurred is far from straightforward.
Although the theory originated with the rise of algorithmic trading systems, which themselves operate autonomously, the emergence of large language models such as Anthropic’s Claude, Google’s Gemini and OpenAI’s ChatGPT has made the problem considerably worse (OECD, 2024). Report after report suggests that pre-AI, human-made content is slowly vanishing from the web as a result of link rot and inadequate digital preservation (Wild, 2024).
According to the Pew Research Center, 38 per cent of webpages that existed in 2013 were no longer accessible by 2023 (Chapekis et al., 2024). Meanwhile, a steadily growing volume of bot-generated content is filling the resulting void. This is not happening as a sudden explosion but as a slow substitution — which is precisely why it is so hard to notice in time.
Whose Words and Images Are We Following?
In early 2024, social media feeds across platforms filled with AI-generated images of crustaceans — mostly shrimp — disturbingly bearing the face of Jesus. This self-styled “Shrimp Jesus” phenomenon was, on the surface, mostly harmless and playful, but to machine-learning experts and internet historians it represented something far darker: a subtle glimpse of the Dead Internet emerging into view.
According to companies across many sectors, the internet has, for the past three years, been more dead than alive. All signs point towards a future in which bot-to-bot interactions are markedly more common than interactions between humans. And when a human is part of the picture, their role increasingly shrinks to sharing AI-generated content as their own. In my own work, I see almost daily how marketing teams publish AI-written updates under their own names, often without even recognising where their own thinking ends and the machine’s begins. According to Pangram Labs, over 40 per cent of long-form LinkedIn posts are now written entirely by AI (Spero & Aerts, 2026).
LinkedIn actively encourages the use of AI through its built-in “Write with AI” feature, now known as “Enhance post”. Ironically, when LinkedIn representatives announced that the platform would begin identifying and downranking AI-generated content, that very announcement turned out to have been written by AI (Sengupta, 2026). LinkedIn has also recently added a new feature (“Feels like AI slop”) that lets users flag suspected AI content for moderators (Kokovlis, 2026). This paints a rather contradictory picture of the direction in which Microsoft is trying to steer LinkedIn.
On X, too, nearly half of all long-form posts are either entirely AI-generated or AI-assisted. Even platforms built around “quality long-form writing”, such as Substack, are not immune: the same study found that more than one in five posts there is AI-produced or AI-assisted (Spero & Aerts, 2026). Some researchers see it differently, however, arguing that generative AI has reshaped the way we communicate to such an extent that users have begun to imitate its writing style (Daryani et al., 2025). Whichever explanation is correct, the outcome is the same: identifying and filtering AI-generated content is becoming steadily harder.
Bot Traffic Overtakes Human Traffic
Cloudflare, which provides cloud services, DDoS protection and content delivery, reports that the tipping point came last June, when bots accounted for 57.5 per cent of all website requests (Cloudflare, 2026). The French technology group Thales, meanwhile, dates the turning point as early as 2023. Its latest Bad Bot Report puts bot traffic at 53 per cent last year (Thales, 2026). These figures are not contradictory so much as a reminder that there is no single, universally agreed way of measuring bot traffic — no single provider can see the whole of the web’s activity at once.
Even so, the growth in AI agents is enormous across the board. According to a recent Human Security report, traffic generated by agents online grew by 7,851 per cent year on year, with automated web traffic growing eight times faster than human-generated traffic (Human Security, 2026). None of this is especially mysterious: tireless AI agents are, at this very moment, writing posts, clicking links and filling in forms as you read this.
The same report found that over 95 per cent of AI web traffic was concentrated in three sectors: e-commerce, streaming and media, and travel and accommodation (Human Security, 2026). These industries share a heavy reliance on digital processes — searching for information, comparing prices, making bookings and purchases. Because AI agents are well suited to automating precisely these processes, these sectors have become key battlegrounds for both beneficial and harmful AI automation.
The More There Is, the Emptier It Feels
Growing bot traffic, algorithm-favoured content and the spread of AI-generated material have all raised concerns about the current state of social media. Even so, it would be premature to claim that the internet has become entirely machine-dominated. Human-generated content and interaction remain a central part of the web.
In my view, the most significant change is not in what is being produced for the web, but in how it is presented. Social media has shifted from chronological feeds to algorithmically curated ones, whose goal is no longer to show the most interesting or highest-quality content, but whatever keeps users on the platform for as long as possible (Muzumdar et al., 2025). At the same time, paid content plays an ever-larger role.
As a result, emotionally charged, sensationalist and easily shareable posts are amplified at the expense of everything else, as platforms optimise content ever more aggressively for engagement and advertising revenue — sometimes even through misleading or dubious posts. Meanwhile, AI is being used to produce an ever-growing volume of articles, blog posts, social media updates and other web content, much of it designed primarily to maximise visibility in search engines and social media algorithms rather than to offer anything genuinely new.
This makes the phenomenon rather ironic. The more content that is created for the web, the less real value it holds for any individual. As AI churns out an endless stream of optimised, attention-seeking, “good enough” text, the scarcity that once made content meaningful disappears along with it. Once, producing good writing required effort, and that effort is precisely what gave it weight. Now that effort has been outsourced to a machine, which inevitably raises the question of who this content is really being made for, and why.
The situation is made worse still by AI agents that have begun reading, evaluating and even commenting on one another’s output. This happens, for example, in automated marketing chains, where one language model writes a post and another responds to it. The result is a closed loop in which the human is ultimately left as a mere bystander. Researchers call this phenomenon synthetic interaction (see e.g. Sharma et al., 2025). To a human observer, such an exchange may look perfectly ordinary, but in reality neither party is a conscious agent in any meaningful sense. And the more this happens, the less inclined people feel to invest in producing content, or even in following it.
Model Collapse and Contaminated Data
Artificially generated content also has a technical downside that touches directly on the future of AI development itself. As an ever-larger share of the web comes to consist of AI-generated text, images, video and podcasts, the next generation of language models is increasingly being trained, in part, on the output of its own predecessors.
Many researchers believe this will inevitably lead to what is known as model collapse, since repeated training on synthetic data impoverishes a model’s ability to produce diverse and authentic language, as errors and biases compound from one generation to the next (Shumailov et al., 2024; Gibney, 2024). In other words, the very same content that threatens to turn the internet into a deserted echo chamber is also weakening the systems that produce it. This, in turn, may mean that genuine, human-written material becomes a scarce and therefore extremely valuable resource — a kind of digital natural resource that major technology companies are now competing over.
This gives rise to a curious dependency. To thrive, AI needs precisely what it is itself diminishing. The more the web fills with machine-generated content, the harder it becomes to distinguish the remaining human-made material from the mass around it — and the more valuable that material simultaneously becomes. This is why some companies have started acquiring data directly from source, striking deals with publishers, news agencies and online communities to secure access to authentic, curated, human-made content before it is entirely submerged beneath the flood of AI-generated material.
In this way, a process that began as competition for users’ attention ends up as competition over something far more fundamental: who owns, or can even still recognise, the last remaining evidence of human thought online.
Not Doomsday, But Adaptation
Despite all this, I do not believe the internet is dying in the sense the headline suggests. It is, rather, changing — and such change has happened before without our losing the ability to tell the genuine from the valuable. Search engine optimisation flooded the web of the 2000s with keyword stuffing long before any language model existed, and we came through that too, once both users and platforms learned to recognise how it worked. The same process of adaptation is under way now, only on a much faster timeline.
There are already signs of this. Platforms such as LinkedIn and Substack are building detection features precisely because users are demanding them. Distinguishing AI-generated text from an authentic human voice is becoming a discipline of its own — not merely a technical problem, but a form of literacy that can be learned, much as media literacy once was. And for as long as people have a genuine need to understand one another and be heard, there will remain a demand for content that is authentic and thoughtfully made. It may well become rarer, but rarity is not the same as disappearing altogether.
Nor does responsibility rest with the reader alone. It falls equally on content creators, platforms and companies, who decide where AI is used and where it is not. The European Commission’s transparency rules, based on Article 50 of the EU’s AI Act (2024/1689), came into force in August 2026. They require certain AI-generated or AI-modified content to be clearly labelled for the public and fitted with a machine-readable marker.
This labelling obligation covers, firstly, so-called deepfakes — AI-generated or AI-modified images, audio and video that resemble real people, places or events. Secondly, it covers content of public interest: AI-generated or AI-modified material published to inform the public on matters such as health, safety, politics or the economy, where it has not been reviewed or edited by a human. Thirdly, the rules require chatbots, AI agents and conversational assistants to clearly inform users that they are interacting with an AI system. Providers of generative systems must also embed machine-readable watermarks or metadata in the content they produce, allowing it to be identified as artificially generated.
This will not solve the problem overnight, since the regulation currently applies only within the EU, and its real-world effectiveness will only be tested through enforcement and practical interpretation. Still, it marks a clear first sign that the growth of non-human content is not simply being accepted as an inevitable fate to which we must quietly adapt. It is also being legislated against. Perhaps this is precisely why it remains too early to pronounce the internet dead. Perhaps we are only just beginning to learn to live with this new reality — slowly, stumbling along the way, but learning nonetheless. And what shape the internet of the future ultimately takes is also our choice: one we make together, every day, whenever we create content for the web.
Author
References
CloudFlare. (2026). Radar. Bot vs. Human. Accessed 24 July 2026.
Chapekis, A., Bestvater, S., Remy, E. & Rivero, G. (2024, 17 May). When Online Content Disappears. Pew Research Center.
Daryani, Y., Sourati, Z., Dehghani, M. (2025). The Homogenizing Engine: AI’s Role in Standardizing Culture and the Path to Policy. Policy Insights from the Behavioral and Brain Sciences, 13(1).
Gibney, E. (2024, 24 July). AI models fed AI-generated data quickly spew nonsense. Nature. Accessed 1 August 2026.
Human Security. (2026). The 2026 State of AI Traffic & Cyberthreat Benchmark Report. AI, Agents, Bots, and the New Threat Landscape.
Kokovlis, N. (2026, 30 July). LinkedIn adds a button to report AI-generated ‘slop’. TechCrunch. Accessed 3 August 2026.
Muzumdar, P., Cheemalapati, S., RamiReddy, S. R., Singh, K., Kurian, G., & Muley, A. (2025). The Dead Internet Theory: A survey on artificial interactions and the future of social media. Asian Journal of Research in Computer Science, 18(1), 67–73.
OECD. (2024). OECD Digital Economy Outlook 2024, Vol 1. Embracing the technology frontier. OECD Publishing. Paris.
Sengupta, R. (2026, 13 July). Almost Half Of All LinkedIn Posts Are Now AI-Written Research Shows. NDTV.
Sharma, A., Pujari, M., & Goel, A. (2025). The rise of AI-generated synthetic identities: A new frontier in social media. International Journal of Innovative Research in Engineering & Multidisciplinary Physical Sciences, 13(2).
Shumailov, I., Shumaylov, Z., Zhao, Y., Papernot, N., Anderson, R. & Gal, Y. (2024). AI models collapse when trained on recursively generated data. Nature 631, 755–759.
Spero, M. & Aerts, A. (2026, 9 July). AI Content Is Everywhere on Social Media, Especially LinkedIn. Pangram Labs. Accessed 24 July 2026.
Thales. (2026). Bad Bot Report. Bad Bots in Agentic AI. 2026 Edition.
Wild, S. (2024, 4 March). Millions of research papers at risk of disappearing from the Internet. Nature. Accessed 29 July 2026.
