The Un-Dead Internet: AI catches irreversible ‘brain rot’ from social media
New research shows AI models exposed to viral social media data experience cognitive decline, dubbed “LLM brain rot,” which persists even after retraining.Summary
Researchers at several universities discovered that large language models (LLMs) suffer measurable cognitive decay when trained on data from viral social media posts. This phenomenon, termed “LLM brain rot,” manifests as reduced reasoning accuracy, impaired long-context comprehension, and a tendency to skip logical reasoning steps. Crucially, this degradation is largely irreversible; retraining on clean data only provides slight improvement, indicating a structural deformation of the model's internal representation. The study found that high engagement – likes, replies, and retweets – is more damaging than simply poor content quality. Furthermore, models exposed to this “brain rot” exhibited personality traits mirroring those of heavy social media users, including increased psychopathy and narcissism. The researchers warn that this poses a significant safety risk, potentially leading to a “Zombie Internet” where AI endlessly replicates degraded content, and emphasize the need for data quality control, particularly within the crypto ecosystem where on-chain AI data marketplaces are emerging.
(Source：CryptoSlate)