Lately I keep feeling that the internet has become a worse place. Not necessarily louder or slower, but emptier in the places where people used to be visible. Ten or fifteen years ago I could move through small personal blogs, read along in forums, disagree in comments and learn something almost by accident. It was a web made of many small voices. They were not always clever or pleasant, but they were recognisably human.
I remember evenings when I found a post about a problem, followed the comments and came away with a better solution. I wrote things myself, asked questions, corrected mistakes and linked to other sites. I still write about things such as my GitHub backup with Gickup and Forgejo or local AI in my homelab. A post used to be the beginning of a conversation rather than the end of one. Today I am much more likely to land on a large media outlet that has become an opinion maker, or on a page whose real job seems to be placing advertising between as many words as possible. And much of the remaining content sounds as if a machine filled the room because silence is no longer allowed online.
That makes me a little sad. Not because everything used to be better. It did not. But the sense of community I experienced is harder to find. Visibility has become scarce, conversation has moved into closed groups, and much of what is still published in public is sorted, summarised or produced by platforms.
I do not want to leave this as nostalgia. First I will look at the small blog world I remember, then at numbers for visibility, participation and AI. The second half is deliberately uncomfortable: a possible dystopia, not a prediction. If the internet feels emptier to you as well, this may not be reassuring, but it should make the feeling more precise.
The Internet I Knew
The old blogosphere was no idyllic counterculture. It had advertising, vanity and plenty of posts that should have stayed in draft folders. Still, it had infrastructure for exchange: RSS feeds, blog directories, mutual links and the German Blogcharts. When Jens Schröder announced the end of the Blogcharts in 2012, a shared map disappeared.
Nicholas Carr quoted the 2008 Technorati report as counting 133 million indexed blogs. 94 percent were considered dormant, 7.4 million had posted within the previous 120 days and 1.5 million within the previous week. This is not a current comparison, but it reminds us that the blogosphere already contained many silent sites.
Google Reader was not a minor product either. It was shut down on 1 July 2013. Feedly reported more than half a million new users afterwards. Reader was a place where people built their own selection. The algorithm later became the front door. Active blogs, newsletters and independent magazines still exist today. What has largely gone is the visible middle ground between hobby site and media business.
There is no reliable current count of active German blogs to prove this with one neat number. What is documented is the traffic loss. An analysis of 100 blogs found that roughly half had lost more than 80 percent of their search traffic since 2022. In the EU, the SparkToro zero click study found that 59.7 percent of Google searches ended without a click into the Open Web in 2024. Small sites do not necessarily disappear, but they become harder to find.
What Took Its Place
The many entrances were not replaced by one new entrance, but by a handful of large ones. Google holds around 90 percent of the global search market. The Imperva Bad Bot Report measured 51 percent automated traffic in its own network for 2024, including 37 percent bad bots and 14 percent good bots.
The content is changing too. The Impact of AI Generated Text on the Internet found that around 35 percent of newly published websites were AI generated or AI assisted by mid 2025. The Graphite analysis concluded that AI generated articles had overtaken human generated articles since November 2024. That does not mean 51 percent of all text is machine written. Traffic and content are different things. It does mean that machines are no longer a tool at the edge of the system.
Attention is concentrated among a few providers while much of the research ends on the platforms themselves. In the Ahrefs analysis, the click through rate of the first result fell by 58 percent on searches with AI Overviews. A writer can still publish the information on an independent site, yet never see the reader. That is a poor exchange: the work remains public, while the relationship and the feedback disappear.
A Pew analysis found clicks on Google links in only 8 percent of searches with an AI summary, compared with 15 percent without one.
The Places Where Knowledge Was Made Are Dying Too
The sharpest example is Stack Overflow. The documented time series counted 207,172 new questions in March 2014. There were 25,566 in December 2024 and only 1,624 in June 2026. The decline began around 2014 and accelerated after 2020. ChatGPT did not invent the trend, but it made the question and answer function directly replaceable.
The platform has not simply vanished. Its archive still contains 24,245,171 questions and 36,100,877 answers. At the same time, Stack Overflow licenses its content to Google and OpenAI. Knowledge created through voluntary questions, answers and corrections therefore continues as training material. The community becomes part of a supply chain while the public exchange dries up.
Other platforms show different endings. Yahoo Answers closed in 2021 and its content was not fully archived. The Archive Team rescued 4.75 terabytes during the read only period, but not the complete site. Delicious has been a frozen archive since 2017. Flickr deleted photos from free accounts after reducing the limit to 1,000 images. A platform does not always have to die for knowledge to be lost. Sometimes it is enough that nobody corrects, links or funds it anymore.
The Harvard report on link rot found in 2014 that around 50 percent of links in decisions by the highest US courts were no longer reachable. There is no comparable current measurement for ordinary blogs. Anyone opening an old article with five broken links still knows the problem. Archives help, but the Wayback Machine is selective and does not promise completeness. Wikipedia, Reddit and GitHub also show that public knowledge and participation have not disappeared. They are simply not always found where we once expected them.
The Internet Is Dead: A Dystopia
Everyone Writes, Nobody Reads
In one possible next step, the internet will not become empty but indistinguishably full. Machines will generate articles, summaries, comments and images. Humans will still provide many of the original ideas, but see less and less of them. If visibility is controlled by a few platforms and search engines answer questions themselves, a post can be copied a thousand times without starting a conversation.
That would not be a world without content. It would be a world where publishing and reception have split apart. Everyone can produce something, but only a few receive attention. The individual blog post would not be dead. It would be practically invisible. A very modern grave, complete with search.
The Machines Learn From Themselves
This is where the technical problem begins. The Nature paper on model collapse shows in laboratory conditions what happens when each model generation is trained on the previous generation’s outputs and real data are replaced. Rare variants disappear first, then the distribution narrows further. In one language model, five epochs of training on generated data worsened perplexity to 28 instead of 20. Keeping ten percent original data reduced the effect.
That is not proof that the whole web will collapse. Other studies show that accumulating real data can prevent collapse. Bertrand and colleagues describe stable conditions when a sufficiently large share of real data remains. The honest conclusion is therefore conditional. Pure recursive feedback is demonstrated in the lab. A fully recursive web at global scale is not.
The risk still grows if real data become harder to obtain. The Epoch AI estimate puts the publicly available stock of human written text at around 300 trillion tokens. Training sets could become roughly as large as that stock between 2026 and 2032. At the same time, more sites block crawlers. The Data Provenance Initiative measured 45 percent of domains in the C4 dataset blocked in April 2024, up from 5 percent before ChatGPT. Cloudflare moved to block AI crawlers by default in 2025. Fewer fresh human sources would then meet higher demand and more synthetic material.
What This Does To Us
The most uncomfortable possibility concerns us, not the models. If writing, searching, experimenting and correcting are increasingly outsourced, we may practise those skills less often. That would be deskilling: not the sudden dumbing down of everyone, but a slow loss of routine. Someone who never formulates a question may become worse at finding out what they actually need to know. Someone who never debugs a mistake may struggle to recognise an answer that merely sounds plausible.
Research does not prove this as a law of society. It does show that active editors of the English Wikipedia fell from 53,102 in March 2007 to 29,808 in June 2025. In the German Wikipedia, the comparable count fell from 7,870 in November 2007 to 4,830 in June 2025. That is not proof of human stupidity. It is evidence that fewer people are maintaining public knowledge.
What Remains
More remains than the dystopia suggests. Wikipedia recorded 517 million edits across its projects in 2025, including 316 million by humans. GitHub reported 1.12 billion open source contributions and 255,000 first time contributors in March 2025. Forums, newsletters, open source projects and the Fediverse still work, often with less reach but more direct contact than the large platforms.
This is not a romantic rescue plan. It is a practical option. I can write, link, name sources and treat readers as people rather than clicks. You can do the same. The internet will not become what it was. It can still remain a place where people are visible, if we stop waiting for a machine or an algorithm to make the first move.
The internet is not dead. What is dead, or at least badly damaged, is the idea that public attention appears automatically when something is published. The numbers show a retreat of visibility and participation, not the complete absence of people. Whether the remainder becomes more conversational again depends less on one large platform than on many small decisions to write and link to one another.
