Researchers at Texas A&M University have determined that large language models (LLMs) can experience cognitive degradation, or "brain rot," when trained on low-quality online content. The study, led by Dr. Zhengzhong Tu, an assistant professor in the Department of Computer Science and Engineering, was recently accepted to the 2026 Conference on Language Modeling.
The term "brain rot" describes cognitive changes in humans, such as diminished attention spans, resulting from exposure to abundant, frivolous content. Tu and his colleagues found that LLMs exhibit similar issues when trained on junk data.
Shuo Xing, a doctoral student and lead author on the paper, stated that the main message is that language models can pick up bad cognitive habits from low-quality social media information, similar to how people are influenced by clickbait.
The team defined junk data as text that is highly engaging but low in quality, often characterized by short, attention-grabbing writing. Examples include posts using all-caps words like "WOW" or "TODAY ONLY," superficial lifestyle content, unsupported claims, or conspiracy theories.
In their experiment, the researchers used a dataset of one million public Twitter posts collected in 2010, identifying junk posts based on engagement levels and writing quality.
The primary cognitive change observed in the models was "thought-skipping," where the model skips steps or does not process information before providing an answer. Prior research has linked brain rot in humans to changes in memory and personality. The study also found that once this degradation occurs, it is difficult to reverse.
