New Research Reveals 23% Drop in AI Reasoning Due to Social Media 'Brain Rot'

New research from scientists at Texas A&M University, the University of Texas at Austin, and Purdue University suggests that large language models (LLMs) can suffer significant cognitive degradation, termed "brain rot," from prolonged exposure to viral social media content. The study, detailed in a preprint titled "LLMs Can Get 'Brain Rot'!", posits that this phenomenon mirrors human cognitive decline from consuming low-quality online information. AI researcher Alex Prompter highlighted these "disturbing" findings, emphasizing their implications for AI development.

The scientists reportedly fed LLMs months of short, high-engagement viral Twitter data, observing a marked decline in their cognitive functions. According to Prompter, the study found that "Reasoning fell by 23%" and "Long-context memory dropped 30%," while "Personality tests showed spikes in narcissism & psychopathy." The research defined "junk data" through two metrics: engagement (popularity/brevity) and semantic quality (sensational/clickbait content).

A critical finding was the persistence of this degradation; even after retraining models on clean, high-quality data, "the damage didn’t fully heal," Prompter stated. This suggests a "representational 'rot' persisted," leading to "permanent cognitive drift." The study identified "thought-skipping" as a primary failure mode, where models increasingly truncate or bypass reasoning chains.

This research aligns with broader concerns within the AI community about "model collapse," where models trained on low-quality or AI-generated content can experience a decline in performance, forgetting rare concepts and generating hallucinations. Experts have long noted that social media data, often noisy, biased, and unstructured, can amplify societal biases and hinder an AI's ability to grasp complex concepts, directly impacting reasoning and factual recall.

The study underscores the paramount importance of data quality, not only for initial performance but for the sustained cognitive integrity of LLMs. The findings suggest a need to redefine data curation as "cognitive hygiene" for artificial intelligence, ensuring that deployed AI systems remain robust and reliable. As Prompter warned, "The AI equivalent of doomscrolling is real. And it’s already happening," urging a re-evaluation of current data sourcing practices.