Researchers have proposed a new hypothesis: that large language models, like humans, can suffer from 'brain rot' after prolonged exposure to low-quality web text. The paper, posted on arXiv, outlines a pilot study designed to test whether junk content causes lasting cognitive decline in LLMs.

The authors created a controlled experiment using real-world data from Twitter/X, aiming to isolate the effects of low-quality text on model behavior. The hypothesis suggests that the degradation is not just a temporary context effect but a lasting change in the model's capabilities.

If confirmed, the findings would have implications for how training and fine-tuning data are curated, particularly as models increasingly ingest unfiltered web content. However, the study is explicitly a pilot, leaving open questions about the scope and permanence of the effect.