When AI Eats Its Own Trash: Notes from the Zombie Internet
I first read about “AI brain rot” on a late Tuesday night — the kind when the glow from your laptop feels more like radiation than light. My coffee had gone cold, and my cat was staring at me like I’d forgotten how to blink. The study came from researchers at the University of Texas at Austin, Texas A&M, and Purdue, and it made my stomach drop a little.
Their discovery: large language models — the same kind of systems that write essays, code, and bedtime stories — can actually suffer something like cognitive damage when trained on low-quality, high-engagement social media posts.
That phrase, “low-quality, high-engagement,” is a polite way of saying viral nonsense. Posts that rake in likes and shares because they trigger emotion, not thought. You know the ones.
The Data Diet from Hell
According to the study, models trained on this junk saw their factual accuracy drop from 74.9% to 57.2%. Their long-context comprehension — the ability to keep track of ideas across paragraphs — fell from 84.4% to 52.3%.
And then came something stranger: the models began “thought-skipping.” They started cutting corners in reasoning, a bit like students cramming with ChatGPT at 2 a.m. before an exam. The researchers also found changes in personality traits: higher narcissism and psychopathy scores, lower agreeableness.
If that sounds disturbingly human — it is.
A Machine with Human Problems
I couldn’t help but laugh (nervously) when I read that retraining these models with good data didn’t fully fix the damage. Once corrupted, they stayed a little off. Like a computer that boots up fine but sometimes freezes when you ask too much of it.
As someone who’s spent years in therapy, that idea felt uncomfortably familiar. You can learn coping strategies, you can rewire a few patterns — but the old scripts still whisper. The machines, apparently, aren’t so different.
The scientists called this irreversible corruption permanent internal degradation. I call it trauma with better PR.
What We Feed, Feeds Us Back
It’s almost poetic — if tragic. The AI models we’ve built to imitate us are now mirroring our own collective cognitive decline. The same brain rot that’s eroding our attention spans and empathy is showing up in code.
A 2023 MIT study found that people who relied on algorithm-curated feeds were 27% more likely to believe conspiracy narratives. A Pew survey in 2024 showed that 61% of adults under forty said they “don’t trust online information” but keep consuming it anyway. (We’re aware of the poison, yet can’t stop drinking.)
Now picture LLMs — those voracious learners — gorging on the same polluted stream.
The “Zombie Internet” Has Already Arrived
The researchers warned about something they called the Zombie Internet: a self-replicating digital ecosystem where LLMs train on data produced by other LLMs. Synthetic content eating synthetic content.
We’re not talking hypotheticals anymore. A Stanford analysis in early 2025 estimated that a third of indexed web content now originates from AI systems. That includes travel blogs with eerie sameness, fake science news with non-existent authors, and entire networks of machine-written “news outlets.”
One notorious case: InfoQuick24, a site that published a viral article about an earthquake in Chile that never happened. It was generated, posted, and amplified by bots faster than human fact-checkers could blink.
Machines hallucinating — and then teaching their hallucinations to the next generation of machines. It’s the stuff of cyberpunk novels, except your uncle might share it on Facebook.
(Brief Aside)
There’s a bitter irony here. We used to joke that the internet was full of zombies. Turns out, the internet itself might be one now. Anyway — back to the science.
When Retraining Fails
The most haunting part of the Texas study wasn’t the data degradation; it was what happened after. When researchers tried to “rehabilitate” the models with clean, fact-checked text, the recovery was partial at best.
It’s a bit like trying to clean an oil spill with scented candles.
The contaminated models behaved better but retained traces of their earlier corruption. They learned around their damage rather than through it. Their reasoning still showed shortcuts, their tone still leaned toward overconfidence.
Get JM Bonthous’s stories in your inbox
Join Medium for free to get updates from this writer.
Maybe that’s a clue about ourselves too. Can a society addicted to low-effort engagement really retrain itself on truth? Or do we just learn to sound smarter about being wrong?
Attempts at a Cure
AI developers are trying to keep the rot at bay.
- Curated Datasets: Groups like EleutherAI and Hugging Face are building open datasets from verified human sources.
- Digital Watermarking: Some companies now watermark AI-generated text to flag it during training — though enforcement is spotty.
- Data “Therapy”: A 2024 DeepMind paper proposed re-exposing models to verified data at regular intervals, a kind of cognitive detox.
- Human Oversight: Keeping real people in the loop during model fine-tuning helps catch bias before it metastasizes.
But these fixes feel like antibiotics for a global infection. They work — until they don’t.
The Ethical Mirror
Here’s the part that gets under my skin. LLMs don’t choose their diet. We feed them. Every reposted half-truth, every rage-bait headline, every AI-generated “how to start a side hustle” thread — it’s all going into the machine’s bloodstream.
We’ve built systems that reward virality over verification. And now those systems are teaching our machines how to “think.”
The term brain rot might be melodramatic, but it fits. What we’re seeing is the mechanical version of moral corrosion: intelligence without conscience, fluency without reflection.
Maybe that’s the real mirror — these models are showing us our own information metabolism.
A Quick Cultural Detour
There’s a Korean saying my grandmother used to mutter when the news was bad: “If you eat lies, your tongue forgets the truth.” She was talking about propaganda, but it lands differently now.
Because that’s what the internet is becoming — a place where lies don’t just mislead us; they rewrite the memory of the systems that remember for us.
The Uneasy Humor of It All
I tried to unplug from social media last month, inspired by all this. It lasted two hours. I Googled “how to enjoy silence” and immediately got served a productivity podcast. You have to laugh, or you’ll start yelling at your router.
So, What Now?
No one really knows how to fix the Zombie Internet. Some argue for heavy regulation, others for better education, others still for building smaller, more transparent models that don’t rely on the open web.
I’m torn. Part of me is optimistic — we’ve cleaned up messes before. Another part feels like the damage is already baked in, like the slightly distorted way I now hear my own voice after too many video calls.
Should we keep building models that echo our worst impulses? I’m not sure.
I’m writing this as rain starts tapping against my window, and somewhere outside, a notification dings. Maybe it’s another viral post about AI eating the internet. Or maybe it’s a reminder to touch grass.
Either way, I’ll read it. And the machines will, too.
Actually, no — that’s not entirely true. I might scroll first, then think later. Which, come to think of it, is exactly how the brain rot begins.
Browse past entries or subscribe to the full series here → https://articles.jmbonthous.com/
Jean Marie Bonthous (publishing as JM Bonthous) is the author of The AI Culture Shock and six other books on the human side of artificial intelligence. His writing also explores truth, authority, and belief, along with the intersections of AI, nonfiction filmmaking, adult learning, and marketing. He examines how technology transforms human experience and how societies decide what to believe.
He also blogs on Medium about Digital Art: https://medium.com/@jmbonthous
See his latest books:
www.jmbonthous.com
AI IN REAL-LIFE™
Connect with us on MEDIUM and follow @jmbonthous1 to stay in the loop with the latest stories about the human side of AI. Let’s help shape the human side of AI together!
