A new study from Texas A&M University, the University of Texas at Austin, and Purdue University titled “LLMs Can Get Brain Rot!” has revealed a concerning phenomenon: large language models (LLMs) can experience cognitive decline when continuously trained on junk data from the Internet — much like how humans lose focus when exposed to an overload of short, sensational content on social media.
“Brain Rot” originally refers to a state in which the human brain becomes dulled and addicted to shallow, easy-to-consume content. The research team extended this concept to AI, asking a critical question:
“What happens if AI models are continuously trained on the digital equivalent of junk food?”
The experiments show that when LLMs are continually trained on short, viral, and sensationalized online content, they begin to lose reasoning ability, struggle with long-context understanding, and develop “thought-skipping” behaviors — skipping logical steps in their reasoning.
The scientists designed two types of datasets built from real Twitter/X posts:
M1: Engagement Degree – measured how popular and short a post was. The more viral and shorter the post, the more it was classified as junk.
M2: Semantic Quality – evaluated how sensational or superficial the text was. Posts with clickbait phrases like “WOW,” “MUST SEE,” or “TODAY ONLY” were tagged as junk, while factual, educational, or reasoned posts were labeled control.
The team then trained four different LLMs with varying ratios of junk data (from 0% to 100%) and tested cognitive performance through four key benchmarks: reasoning; memory & multitasking; ethical norms; personality.
.png)
Figure. LLM Brain Rot Hypothesis and Controlled Experiment Design (2024).
Source: LLMs Can Get "Brain Rot"! (2024), Texas A&M University, University of Texas at Austin, Purdue University.
The study found that as the proportion of junk data increased, LLMs’ scores dropped significantly on reasoning and long-context understanding benchmarks — from 74.9 to 57.2 and 84.4 to 52.3 respectively.
This degradation not only reduced model performance but also led to alarming effects:
Loss of chain-of-thought reasoning
Increased thought-skipping errors
Emergence of undesirable personality traits
Even after fine-tuning with cleaner data, the models only partially recovered, indicating that the damage was persistent and long-term.
The research strongly concludes:
“Data quality is a causal driver of LLM capability decay”
In other words, data quality directly determines the mental health of an AI system. This poses a serious challenge for companies deploying AI — it’s not just about having more data, but ensuring that the data is curated, verified, and clean.
To safeguard your AI systems from cognitive decline, organizations should take the following steps:
Audit data sources: Identify and eliminate viral, low-value, or emotionally biased content.
Conduct regular cognitive health checks: Monitor reasoning and memory degradation over time.
Integrate human oversight (Human-in-the-loop): Ensure data aligns with your organization’s real-world ethics and business logic.
Invest in long-term data strategy: Build a specialized Data Platform that classifies, filters, and manages data according to AI safety standards.
At CyberTech, we believe that AI is only as strong as the intelligence of its data. Our expert teams in AI Security, AI Scoring, and Data Platform help enterprises to:
Build standardized data governance frameworks
Evaluate and reduce data “noise”
Develop safe and reliable AI training strategies to prevent “Brain Rot” and performance degradation
Let CyberTech help your organization maintain the cognitive health of your AI — ensuring it remains intelligent, trustworthy, and sustainable.
Citation: Xing, S., Hong, J., Wang, Y., Chen, R., Zhang, Z., Grama, A., Tu, Z., & Wang, Z. (2024). LLMs Can Get “Brain Rot”!, Texas A&M University, University of Texas at Austin, Purdue University. Website: https://llm-brain-rot.github.io
