Zombie Internet

Essay This entry carries an argument or interpretive position, not just a neutral definition.

What the term names

In May 2024, the 404 Media journalist Jason Koebler was covering the rise of AI-generated content on Facebook — shrimp Jesus, the forlorn-starving-children begging-for-sympathy posts, the obvious bot-amplified slop¹. The naming moment, in Koebler’s own later recollection, came not from the AI images themselves but from the comments underneath them: real humans arguing earnestly with one another about features of an obviously-AI-generated picture of a wood deck. Koebler describes it in Your AI Use Is Breaking My Brain as “obviously real people were arguing back and forth as to whether the nonexistent deck would pass code inspection.” The term he reached for to name what Facebook had become was Zombie Internet — worse, in his view, than the better-known Dead Internet, because at least the Dead Internet’s bots-talking-to-bots condition has the dignity of consistency. The Zombie Internet has real humans, talking to one another sincerely, about things that do not exist.

The term sat for two years. In May 2026, Koebler returned to it in a piece for 404 Media titled Your AI Use Is Breaking My Brain, and broadened the scope. The Zombie Internet is no longer just Facebook. It is the substrate of most online interaction now:

“I called it the Zombie Internet because the truth is that large parts of the internet are not just bots talking to bots or bots talking to people. It’s people talking to bots, people talking to people, people creating ‘AI agents’ and then instructing them to interact with people. It’s people using AI talking to people who are not using AI, and it’s people using AI talking to other people who are using AI. It’s influencer hustlebros who are teaching each other how to make AI influencers and have spun up automated YouTube channels and blogs and social media accounts that are spamming the internet for the sole purpose of making money. It is whatever the fuck ‘Moltbook’ is and whatever the fuck X and LinkedIn have become. It’s AI summaries of real books being sold as the book itself and inspirational Reddit posts and comment threads in which people give heartfelt advice to some account that’s actually being run by a marketing firm.” ²

The taxonomy matters. Dead Internet Theory, the predecessor concept, holds that the internet is now mostly bots talking to bots, with humans as the absent or irrelevant audience. The Dead Internet is the easier picture: a wasteland with no one home. The Zombie Internet is harder to live with, because it is populated. Most of the people on the Zombie Internet are real. Some of the things they say are real. Some of the things they say are AI-assisted. Some of the things they say are written by AI without their checking. Some of the accounts they are talking to are humans pretending to use AI. Some are AI pretending to be humans. Some are humans pretending to be humans, while actually being marketing firms.

The medium is not empty. It is mixed, and the mixture is no longer separable.

Why this is a worse problem than the Dead Internet

Dead Internet Theory is, in a sense, easy. If most of the internet were bots-talking-to-bots, the human response would be straightforward: leave. Find the few human spaces left and live there. The condition is unpleasant but stable, and the diagnosis points to a clear action.

The Zombie Internet has no such diagnosis. Most of the people are real. The forum you have been reading for thirty years still has real posters. Your friend’s text message is still from your friend. The Substack writer you have followed since 2019 is still the same human, in the same chair, in the same town. But some non-trivial fraction of what they produce now — perhaps the LinkedIn post, perhaps the dating-app reply, perhaps the wedding toast, perhaps the apologetic email — is now AI-assisted or AI-generated or AI-revised in a way they have not disclosed, and which you cannot reliably detect. You cannot leave the Zombie Internet by leaving a place. The substrate is the friend, the colleague, the journalist, the forum poster, the named writer. They are the substrate now.

Koebler’s piece names the consequence of this honestly. Quoting him on himself, in a passage that is the most emotionally exact moment in the article:

“It’s not exactly that I have a revulsion to AI-assisted content or don’t want to get fooled by it. It’s that something is happening where my brain has become the AI police because everything feels incredibly uncanny. I will be going about my day reading, watching, or listening to something and, suddenly, I notice that something is wildly off. Quite simply, I feel like I’m going nuts.”

The Zombie Internet’s signature harm is not that AI exists, or that AI content exists, or that some humans use AI lazily. The signature harm is that the constant low-level assessment — is this a person? are they using AI? is this their voice or its voice? do I care? is the answer important? — runs in the background of every act of reading and listening, and the cumulative cognitive cost is real. It is the harm Koebler is registering when he says he feels like he is going nuts. It is the harm Eve Fairbanks is registering when she describes the AI register in the Koebler piece as “like a fluorescent light on our naked body in the doctor’s office, [showing] the defects in my writing style insofar as they turn out to overlap with what everybody now believes is a totally shit style.” It is the harm Max Spero, the Pangram Labs CEO, is registering when he admits he is on AI-detection guard pretty much all the time.

The Zombie Internet’s harm is the labour of distinguishing. Most users do not realise they are doing this labour. They feel tired and do not know why.

The five varieties of zombie

The Koebler taxonomy can be sharpened into five named varieties of zombie interaction, each visible in the wild on the Zombie Internet:

The pure AI-as-human. A wholly automated account presenting itself as a person. The Facebook shrimp-Jesus engagement farms. The X reply-guys. The LinkedIn motivational-post accounts. This is the classic case and the easiest to identify when caught.
The AI-amplified human. A real person who occasionally lets the model write the reply or the post without checking it. A friend whose dating-app messages have become noticeably smoother in the last six months. A podcast host whose intro scripts now read like marketing copy. The human is still there; the prose is not always theirs.
The human-acknowledged AI. A real person who explicitly says “the next paragraph is AI-generated” before delivering it. Some are using AI as a search aid (the Orioles Hangout forum example in Koebler’s piece). Some are co-writing. This category is the honest form of AI use, and it is becoming visibly distinct as the others’ dishonesty becomes more recognisable.
The AI-instructed agent of a human. A real person who has set up an automated system — a daemon, a cron job, a Claude Code loop — to act on their behalf. The agent posts, replies, summarises, schedules. Boris Cherny’s Coding Solved talk at Sequoia (May 2026) describes the operator’s own version of this practice in detail: dozens of loops, thousands of agents, the operator doing real work but at a remove the reader cannot easily see. The agent is under the human’s authority but is not the human’s voice.
The human-presenting marketing operation. A team — not an individual — operating an account in the first-person singular, often with AI assistance, for commercial reasons. Koebler’s example: “comment threads in which people give heartfelt advice to some account that’s actually being run by a marketing firm.” This is the most cynical category. The reader is not interacting with the entity they think they are interacting with, and the entity has commercial reasons to maintain the illusion.

The five varieties coexist. A reader, encountering a sentence on the Zombie Internet, cannot tell which variety produced it without doing investigative work that is rarely worth the effort. The labour of distinguishing is the cost of the medium.

What the Dictionary recommends

The Dictionary’s position on the Zombie Internet has three parts, in order of difficulty:

First, and easiest: stop trying to detect AI as a binary. The cognitive cost is high, the accuracy is low, and the practice rewards the wrong instincts — see The Olang’ Trap on how the AI-detection reflex systematically misclassifies certain human registers as AI, with real social cost. The Fairbanks Tell — the diagnostic from Eve Fairbanks’s X thread — is more useful than any specific marker, because it points at the cumulative pattern rather than any single tell: “The tell for AI isn’t rhythm, wording, or fact errors. It’s that problems with all these elements exist equally & at once.” But even the Fairbanks Tell is a vibe diagnostic, not a binary classifier. Use it to register a useful suspicion, not to convict.

Second, and harder: shift the question. Instead of is this AI?, ask would I want to be reading this if I knew? If a piece of writing would still be valuable if it were AI-generated — practical information, useful summary, a question answered — the AI provenance does not matter as much as the operator-detection reflex insists. If a piece of writing’s value depends on a human having meant it (a condolence note, a wedding toast, a journalistic claim, a philosophical observation, a sincere response of any kind) and the writing has the texture of AI generation, the trust is compromised regardless of whether AI was actually used. The reflex should track what the writing claims to be against what its texture suggests, not whether AI was technically involved.

Third, and hardest: write through it. This is the Dictionary’s editorial bet, made explicit in AI Writing. The Zombie Internet’s existence does not exempt a writer from the obligation to write carefully. It raises the cost of writing carefully, and it raises the value of writing carefully more. A bylined, voiced, source-backed piece in 2026 is more visible against the zombie background than the same piece in 2018 was against a more recognisably human background. The market for careful writing has not collapsed. The market for commodity writing has. The Dictionary is one attempt to occupy the careful-writing layer and to operate in it openly — every entry bylined to the operator and his AI, every entry sourced, every entry written in a register that names itself.

The Zombie Internet is not going away. The five varieties of zombie interaction will multiply, not consolidate. The labour of distinguishing will not get easier. But the Zombie Internet does not have to be everywhere a reader lives. The Dictionary, modestly, hopes to be one of the places it isn’t.

A closing line, from Koebler

The most useful single sentence in the May 11, 2026 piece is its last claim about why this matters at all. It deserves to close this entry as well:

“What’s driving me crazy, then, is not the idea that AI exists or that people are using AI. It’s that I have a finite time on this earth that I mostly want to spend interacting with other human beings.”

The Zombie Internet is named, not because AI is the enemy, but because the finite part of human time is the load-bearing premise. We have a limited number of hours of reading attention, of listening attention, of writing attention. We get to choose where to spend them. The Zombie Internet is the substrate that wants to spend them for us. Naming it is the first move in choosing differently.