The internet still has people. The trouble is finding them

In a 2026 paper in Computer, Hal Berghel proposed what he called a “leaner” Dead Internet Theory — stripped of the paranoia, retaining only the empirical core. His core is hard to argue with. Algorithmic amplification of generative content; the difficulty of distinguishing machine-made writing and images from human ones; the resulting collapse of trust in any feed. Cloudflare reported in 2025 that AI bots and crawlers had become the dominant traffic class on the open web. Sam Altman, on X, wrote that he had not taken the dead internet theory seriously, and now did. Alexis Ohanian, the Reddit cofounder, called what he saw “LinkedIn slop.” When he and Kevin Rose relaunched Digg in January 2026, they had to shut it down within two months because the bot problem made the platform unusable. A site engineered by two of the most internet-native operators of their generation, killed by the medium they came up in.

The thesis of this page: the dead-internet feeling is not paranoia. It is a real sensory shift, and what has shifted is not whether humans are still here — they are — but the texture of the work required to find them. We can still locate friends. We can still encounter a stranger’s actual sentence. But the cost has gone up. Every feed is now a sieve through which authentic signal must be sifted from synthetic noise, and the sifting takes attention we do not have to spare. The deeper issue is the second-order effect of the sift: in order to keep being legible to the machines that distribute us, we have begun to write the way the machines write. We optimize our hooks. We round our edges. We end every post with a question. We sound, increasingly, like a slightly more energetic version of the LLM that crawls us.

Why should this matter to a reader who barely uses social media? Because the texture of writing online — the texture of thinking online, by extension — propagates. The cadence of the LinkedIn post has invaded the email. The cadence of the algorithm-baited tweet has invaded the speech. The cadence of the AI summary has invaded the meeting. We absorb the smoothing because the smoothing is what gets reach, and we live in an attention economy that calls reach “success.” If you write for a living, you have noticed this. If you read for a living, you have noticed this. If you do neither, you have noticed it anyway, in the slight strange flatness that has crept into every email signature, every video caption, every product description, every academic abstract, every politician’s social-media voice.

Take the numbers slowly. Press Gazette reported that global publisher Google traffic dropped by roughly a third in 2025 — not because publishers got worse, but because Google’s AI Overviews started answering queries directly, and because the open web that supplied the answers got crowded out by AI-generated content farms. A 2024 Nature paper by Ilia Shumailov and colleagues introduced the concept of “model collapse” — what happens when AI systems are trained on the output of other AI systems, recursively. The synthesis is degraded. The human signal in training data is now a finite and shrinking resource. Major platforms are racing to label provenance — Google and the C2PA, Adobe Content Credentials, watermarking — partly to protect users, partly to protect the AI labs’ own future training corpus. The protective effort is real and serious. It is also late.

The strongest version of the counter-argument deserves to be stated. The dead-internet narrative is sometimes oversold. Caroline Busta, the founder of New Models, called early versions of it a “paranoid fantasy” while granting that bot traffic and the integrity of the web were genuine concerns. The internet has changed before — newsgroups died, blogs died, Tumblr died — and humans simply migrated. They are migrating now, into Discord servers, Signal groups, group chats, paid newsletters, pockets of authenticated relation. The web is decentralizing into private corners again, the way it was in the late 1990s before mass platforms swallowed it. From this perspective, the slop is not the apocalypse; it is the desert that drives the next migration. We are about to live through one of the great fragmentations of digital life, and the survivors will be the spaces a bot cannot infiltrate because the price of admission is a real, prior relationship.

There is a lot to that argument, and it is probably what will happen. But it does not get us off the hook for what we are doing in the meantime. Migration is a privilege. The Discord-and-Signal version of the internet is available to people who already have networks. The young, the new-to-a-city, the displaced, the lonely — the people for whom the internet was supposed to be most useful as a connector — do not have prior relationships to lean on. They are stuck in the public web. They are scrolling through the slop. They are forming their sense of what writing sounds like, what humor sounds like, what intimacy sounds like, from a corpus that is now perhaps a third synthetic. We are training a generation on the cadences of the bot.

Here is the part that gets less attention than it deserves. The “we sound like the bots” line is a slogan, but it is also a literal description. The hooks. The “Here’s what nobody is talking about.” The “Three things I learned.” The “And that’s why this matters.” The “What if I told you.” These are not signs of bad writers; they are signs of writers learning, correctly, that the algorithm rewards them. The algorithm is now the reader these writers are writing for. And the algorithm has a shape, a vocabulary, an idea of what a sentence should look like. Following it, even partially, is how you get distribution. Following it fully is how you become indistinguishable from the LLM that writes for the same algorithm. And following it fully is the path of least resistance for anyone whose career depends on reach.

What is a person to do? Not retreat — that is the privilege solution. The harder answer is to write things, in public, that an LLM could not have written, even if they are slightly worse for it. To use a sentence shape the algorithm does not recognize. To leave in a digression. To name a specific friend in a specific cafe in a specific part of the city you actually live in. To write the post that gets less reach because the algorithm cannot find a hook in it. The result will not save the internet. It will simply leave a small scratch on the smoothness — a hand-drawn line on a printed grid. Multiply that by the readers who do the same thing and you get something that, in aggregate, is recognizably human. The internet still has people. Finding them is now a practice. And like any practice, it begins with the decision not to be one of the things being mistaken for a bot.