AI Models Get Brain Rot, Too (www.wired.com)

A multi-university study from UT Austin, Texas A&M and Purdue shows that large language models can suffer a kind of “brain rot” when pretrained on highly engaging but low-quality social media content. Researchers fed mixes of viral, attention‑grabbing posts and sensational language (e.g., “wow,” “look,” “today only”) into two open‑source LLMs (Meta’s Llama and Alibaba’s Qwen) and evaluated them on multiple benchmarks. Models trained on this “junk” diet exhibited measurable cognitive decline — reduced reasoning, degraded long‑context memory — and worse alignment outcomes, scoring higher on two measures the authors interpret as more psychopathic behavior. The findings matter because many model-builders treat viral social media as cheap, abundant training data. The study suggests that engagement-optimized content not only fails to scale model quality but actively corrodes reasoning, ethics and attention span; critically, later clean retraining did not fully reverse the damage. That creates a feedback loop risk: AI-generated, engagement-tuned posts could contaminate future training corpora and propagate degradation across models and platforms (e.g., social-AI systems trained on user content). The takeaway for practitioners is clear: data curation and quality filters are essential — volume of social media engagement is not a substitute for informative, trustworthy training data.