Sitemap

|AI|LLM|COGNITIVE DECLINE|SAFETY|

Unplug Your AI: Junk Internet Data Is Rotting LLM Brains

Your Model Is What It Eats: What Happens When LLMs Train on the Worst Parts of the Internet

9 min read1 day ago
Press enter or click to view image in full size
This study introduces the “LLM Brain Rot” hypothesis, showing that continual pretraining on low-quality Twitter/X data causes measurable cognitive decline in large language models. Junk data reduces reasoning, long-context ability, and safety, driven by thought-skipping and persistent representational drift. Results highlight data quality as a key factor in AI reliability and the need for routine model health checks.
image generated by the author using AI

In internet culture, the term refers to the detrimental effect on human cognition of consuming a large volume of online content (especially from social media). Internet addiction appears to have a significant effect on human cognition, especially on attention span (reduced ability to maintain focus when reading and solving problems), memory processes (alterations in how an individual stores, retrieves, and prioritizes knowledge), and social cognition (modification of self-concepts and influencing self-esteem).

Since LLMs learn from enormous amounts of data that have a large amount of internet content, can they also experience brain rot?

Given their training with enormous amounts of tokens, LLMs are exposed to a huge amount of junk data, just like humans. Although LLMs do not have “neurons” and “cerebral cortex” like humans, they have both parameters and attention mechanisms that might analogously be ‘overfitted’ or “distracted” by certain data patterns.

Create an account to read the full story.

The author made this story available to Medium members only.
If you’re new to Medium, create a new account to read this story on us.

Or, continue in mobile web
Already have an account? Sign in
Level Up Coding
Salvatore Raieli

Written by Salvatore Raieli

Senior data scientist | about science, machine learning, and AI. Top writer in Artificial Intelligence

Responses (6)

Write a response

Good article
AI document hygiene is important, you can build models for hygiene checkers by examining problem documents using another AI, once you have an understanding of the wonky logic inserts or mythic framing tricks or whatever you find, that can be made into a functioning safeguard layer for an AI

11

✍️ Beautifully put I wrote something similar recently too!

9

Recommended from Medium

See more recommendations