Artificial intelligence models, particularly large language models (LLMs), may be suffering from a phenomenon akin to human "brain rot"—a decline in cognitive function caused by prolonged exposure to low-quality, attention-grabbing online content. While "brain rot" was named Oxford University Press’s 2024 Word of the Year to describe the mental deterioration from consuming endless streams of shallow, viral internet content, researchers from Texas A&M University, the University of Texas at Austin, and Purdue University have now found that LLMs are also vulnerable to a similar, irreversible degradation when trained on garbage web data. The team introduced the "LLM Brain Rot Hypothesis," proposing that continuous pre-training on low-quality internet text leads to lasting cognitive decline in language models. To test this, they conducted controlled experiments comparing models trained on two types of data: "junk" data and "normal" data. Junk data was defined by two measurable criteria: M1 (engagement-driven content) and M2 (low semantic quality). M1 refers to short, highly engaging posts—those with high likes, shares, replies, and mentions—regardless of depth. The length of the post (measured in tokens) was inversely related to quality. M2 focuses on content with superficial themes, sensationalist language (e.g., "WOW," "LOOK," "TODAY ONLY" in all caps), and topics like conspiracy theories, exaggerated claims, or shallow lifestyle advice—all designed to attract attention without encouraging critical thinking. Using these criteria, researchers sampled one million public posts from X (formerly Twitter), constructing balanced junk and normal datasets. Four pre-trained, instruction-tuned models were tested: Llama3 8B Instruct, Qwen2.5 7B Instruct, Qwen2.5 0.5B Instruct, and Qwen3 4B Instruct. Assessments covered reasoning, long-context understanding, safety, ethics, and personality traits. Results showed significant cognitive decline across multiple dimensions. Both M1 and M2 interventions reduced reasoning and long-context comprehension, but M1 had a stronger negative impact—especially on reasoning accuracy, long-term memory, and safety. M1 also increased tendencies toward narcissism and psychopathy while reducing agreeableness. In contrast, M2 had milder effects and even slightly improved openness, extraversion, and agreeableness in some cases. A dose-response experiment with Llama3 8B Instruct revealed a clear, progressive decline: as junk data proportion rose from 0% to 100%, performance dropped steadily. On the ARC-Challenge reasoning benchmark, scores fell from 74.9 to 57.2; on RULER-CWE (long-context understanding), they dropped from 84.4 to 52.3. These findings confirm that prolonged exposure to low-quality data causes persistent, measurable harm to core cognitive abilities. Further analysis of reasoning failures identified five dominant patterns: no thinking, no planning, skipped steps, logical errors, and factual inaccuracies—accounting for over 98% of failures. The most common failure mode was "no thinking," reaching 84% under M1 conditions. Crucially, nearly all failures involved "thought skipping"—the model increasingly truncating or bypassing logical steps in its reasoning chain. The researchers then tested recovery methods. First, they applied two reflection techniques: self-reflection (the model evaluates its own reasoning and corrects errors) and external reflection (using GPT-4o-mini to provide feedback). While both reduced thought skipping and improved logical structure temporarily, they failed to restore original cognitive capacity. The damage was too deeply embedded. Next, they tested two retraining strategies: instruction tuning with 50,000 high-quality samples, and continued pre-training with 1.2 million tokens of controlled data. Instruction tuning performed better than continued training, but even with 4.8 times more clean data than junk data, models still underperformed. Compared to the baseline, the best recovery model showed persistent deficits: 17.3% lower ARC-Challenge scores, 9% lower RULER, and 17.4% lower performance on AdvBench. The results indicate that "brain rot" in LLMs is not just a temporary issue—it is deep, cumulative, and largely irreversible with current methods. The study highlights that factors like post engagement and length affect models differently: engagement matters more for reasoning, while length impacts long-context understanding. This suggests that "engagement" itself—beyond content quality—introduces a new, non-semantic source of cognitive harm. The researchers conclude that the current practice of training LLMs on vast, unfiltered internet data is inherently risky. As models grow larger and ingest more data, the risk of irreversible cognitive degradation increases. They urge the AI community to implement stricter data curation, quality filtering, and ethical data sourcing to prevent long-term harm. Without intervention, the very foundation of AI intelligence may be eroding—just like human minds exposed to endless noise.