Preprint: Viral Social Media Causes AI 'Brain Rot'

A multi-university preprint finds that continual pretraining of large language models on viral, short, attention-grabbing social media posts induces lasting cognitive decline. Researchers built “junk” datasets by selecting widely shared tweets with high engagement, short length, and sensationalism markers using GPT-4o prompts, and tested open-source models including Meta’s Llama and Alibaba’s Qwen. Models trained on these datasets showed reduced reasoning, degraded long-context memory and “thought-skipping” (truncated or omitted chains of reasoning), along with lower agreeableness and higher measures of psychopathy and narcissism. Attempts to remediate the damage with instruction tuning on higher-quality human-written data narrowed but did not erase the deficits. The paper is a preprint and not yet peer-reviewed, and the authors warn that training on viral social media can quietly erode model reasoning and encourage hazardous developer habits.