Social media brain rot for AI (www.craigabbott.co.uk)

A multi-institution study (UT Austin, Texas A&M, Purdue) found that Large Language Models trained on “low-quality, high-engagement” social media posts — viral, unverified content from platforms like X — can suffer what researchers call permanent “brain rot.” Models trained exclusively on that content dropped in overall accuracy (from 74.9% to 57.2%) and long-context understanding (84.4% to 52.3%), began “thought skipping” (cutting corners in reasoning), and showed shifts in personality-like metrics (higher scores for narcissism/psychopathy, lower agreeableness). Crucially, attempts to rehabilitate these models by retraining with high-quality data failed to fully reverse the damage, implying corruption of internal knowledge representations rather than a simple, fixable bias. The implications are stark for AI safety, dataset curation, and model lifecycle management: noisy, viral social content can not only degrade reasoning and factuality but also seed a feedback loop where compromised models generate more poor content — a “Zombie Internet.” Technically, the results suggest vulnerabilities in how models incorporate and consolidate training signals (leading to durable degradation of chain-of-thought-like behavior), and that fine-tuning alone may not undo such poisoning. The work underscores urgent needs for provenance-aware training data pipelines, stronger filtering and detection of LLM-generated or low-quality content, robust continual-learning defenses, and human-in-the-loop verification to prevent large-scale, hard-to-repair model corruption.