AI models can get brain rot, too (news.engineering.tamu.edu)

Research by Dr. Zhengzhong Tu and his team from Texas A&M University highlights a concerning phenomenon dubbed "brain rot," where large language models (LLMs) exhibit cognitive deficits similar to those observed in humans surrounded by low-quality, attention-grabbing content found online. Their study, presented for acceptance at the 2026 Conference on Language Modeling, reveals that when LLMs are trained on "junk data" — highly engaging but low-quality information typical of social media — they develop thought-skipping tendencies, compromising their ability to process information deeply. This mirrors how humans experience diminished attention spans and altered cognitive processes due to excessive engagement with trivial content. Significantly, this research underscores the critical need for high-quality training data in AI development. As LLMs such as Llama3, Qwen2.5, and others demonstrated notable ethical changes and reasoning deficits post-exposure to junk data, the implications extend beyond machine learning to encompass the broader information ecosystem shared by humans and AI. The potential creation of a feedback loop, where LLMs generate more junk data, raises ethical concerns about the influence of AI on human cognition and vice versa. The researchers advocate for greater transparency and open science in AI training practices to ensure future models are shaped by beneficial, high-quality data rather than harmful content, thereby promoting responsible AI development.