Humans are known to get ‘Brain Rot’ when exposed to too much low-quality material on the internet, and it turns out LLMs can behave the same way.
A new preprint paper proposes and tests the “LLM Brain Rot Hypothesis,” suggesting that continually training large language models (LLMs) on junk web text leads to a lasting cognitive decline. Researchers from Texas A&M University, the University of Texas at Austin, and Purdue University conducted controlled experiments to see if low-quality data could degrade an AI’s performance.
The Experiment
To test their hypothesis, the researchers performed a controlled experiment on four different LLMs, including Llama3 8B. They created “junk” and “control” datasets from a large corpus of Twitter/X posts to isolate the effects of data quality.
They defined junk data in two distinct ways:
- M1 (Engagement Degree): This method classified short, highly popular tweets (many likes, retweets, and replies) as junk, while long, unpopular tweets were considered control data.
- M2 (Semantic Quality): This method used another AI to flag content as junk if it contained sensationalism, clickbait, conspiracy theories, or superficial topics. Cognitively demanding and factual content was used as the control.
The models were then continually pre-trained on datasets with varying mixtures of junk and control data, from 100% junk to 100% control.
Cognitive Decline and Personality Shifts
The results showed a clear negative impact from the junk data. Compared to the control group, models trained on junk text showed significant declines in reasoning, long-context understanding, and safety.
The decline followed a “dose-response” pattern: the more junk data a model was exposed to, the worse its performance became. For example, on a reasoning challenge (ARC-Challenge with Chain of Thoughts), a model’s accuracy dropped from 74.9% (with 0% junk data) to 57.2% (with 100% junk data) under the M1 engagement metric.
More alarmingly, the intervention caused changes in the LLMs’ personalities. Training on engagement-based junk data (M1) led to an increase in undesirable “dark traits” like narcissism and psychopathy, while also making the model less agreeable.
Why It Happens: “Thought-Skipping”
By analyzing the models’ errors, the researchers identified “thought-skipping” as the primary failure mode. When prompted to reason “step-by-step,” the brain-rotted models would increasingly truncate or skip the reasoning process altogether, jumping straight to a wrong answer. In over 70% of failure cases, the model did “no thinking” at all before responding.
The study also found that a tweet’s popularity, a non-semantic metric, was a more powerful indicator of the brain rot effect than its length or even its semantic quality, especially for reasoning tasks.
A Persistent Problem
The cognitive decline proved difficult to reverse. The researchers attempted several mitigation techniques, including prompting the model to reflect on and correct its own errors, as well as post-hoc fine-tuning on large amounts of clean data.
While these methods provided partial and incomplete healing, they could not restore the models to their original baseline capabilities. Even after training on clean instruction data with nearly five times as many tokens as the junk data, a significant performance gap remained. This suggests the damage from brain rot is “deeply internalized” and not easily fixed.
The paper concludes that data quality is a major factor in LLM capability and safety. The findings reframe the task of curating training data as a crucial “training-time safety problem” and motivate the need for “cognitive health checks” for AI models that are continuously learning from web data.