The 'Internet Zombie': When Filtering AI Becomes Exhausting

There's a specific moment many of us will recognise: you open an article, read two paragraphs and something feels off. The text is correct, even fluid, but hollow in a way that's hard to pin down. You close it. Open another. Same thing. After twenty minutes, the fatigue isn't in your eyes but something deeper. Jason Koebler, editor of 404 Media, has put a name to that feeling and written about it with a frankness the ecosystem needed.

Simon Willison picked up the article on his blog on May 11th, calling it an "excellent and furious" piece. It's high praise coming from someone who has spent years documenting the state of the AI tools ecosystem with uncommon rigour.

What the Internet Zombie Is (and How It Differs from the Dead Internet)

The "Dead Internet" concept has been circulating in forums for years: the idea that the web has filled with bots talking to other bots, generating artificial activity that displaces real human interaction. Koebler proposes something different and more unsettling.

The Internet Zombie isn't a dead network but one in an intermediate, ambiguous state, difficult to classify. In his own words:

> "I called it Internet Zombie because the truth is that large parts of the internet aren't just bots talking to bots or bots talking to people. They're people talking to bots, people talking to people, people creating 'AI agents' and then instructing them to interact with people. They're people using AI talking to people who don't use it, and they're people using AI talking to other people who also use it."

The distinction matters. A dead internet is, in a sense, ignorable: if bots are talking to each other, the problem is systemic but doesn't directly confront you. The zombie internet does. You're reading someone real who has delegated part of their expression to a model. Or someone real whose style has begun to resemble, unintentionally, the output of an LLM because they've spent months reading that type of prose.

The Cognitive Cost of Filtering

This is perhaps the most compelling part of Koebler's article and what resonates most among those of us working closely with these tools. The problem isn't just that there's more AI-generated content: it's that detecting it consumes mental energy continuously.

It's not a binary filter that activates once. It's constant vigilance, an additional processing layer that accompanies every read. Is this human? Is someone here? Do I care about the answer for what I need right now? That loop, repeated hundreds of times a day, carries a cost that doesn't appear in any productivity metric.

Koebler adds an uncomfortable nuance: stylistic contagion. Writing generated by language models has recognisable quirks—three-part structure, courteous euphemisms, absence of direct opinion, excessive use of certain connectors—and those quirks are starting to appear in human writing. Not because authors use AI, but because they've been reading it so much that it has reconfigured their stylistic references.

Who This Matters For

The article speaks mainly to general readers fed up with bland content, but it has direct implications for several more specific profiles.

• Editorial and content teams that use AI as an assistant: the risk isn't just publishing mediocre text, but contributing to reader exhaustion and eroding trust in the brand itself.
• Developers of agents and automations with Claude or other models: when an agent interacts on behalf of a person in forums, networks or comments, it's actively participating in building the zombie internet that Koebler describes.
• Anyone managing online communities: moderation and recognition of authentic participation become exponentially more complicated when human and automated signals share the same channel.

A Note on the Moment

Koebler isn't the first to point out this phenomenon, but he's among those articulating it most precisely at this particular moment. The conversation about quality and authenticity of online content has been gaining intensity for months, and this article will reach readers who have already been feeling it without the vocabulary to describe it.

From our side, we've been observing how the same technology that enables useful integrations (MCP servers, agents, legitimate automations) can, misapplied or applied without discernment, feed exactly the problem Koebler documents. The tool isn't the problem; the absence of editorial judgment about when and how to use it is.