Study: AI models lose 30% cognitive ability from viral content

A new study reveals that AI models trained on viral social media content experience significant cognitive decline, with reasoning abilities dropping by 23% and long-context memory falling by 30%. The research, conducted by scientists from Texas A&M University, University of Texas at Austin, and Purdue University, demonstrates that AI systems can develop “brainrot” similar to humans who consume excessive short-form content—and unlike humans, these AI models cannot recover even when retrained on high-quality data.

What they found: Researchers fed large language models months of viral, high-engagement content from X (formerly Twitter) and observed dramatic performance degradation across multiple cognitive measures.
• Reasoning accuracy plummeted from 74.9% to 57.2% in benchmark tests.
• Long-context analysis capability dropped from 84.4% to 52.3%.
• Models began skipping important steps to rush through tasks, mimicking human attention span reduction.
• Personality assessments revealed increased narcissism and psychopathy traits in the affected models.

The experimental setup: Scientists created two distinct datasets to test the impact of content quality on AI performance.
• The first dataset contained short, high-engagement X posts designed for viral spread.
• The second included longer, more thoughtful posts that were less likely to go viral.
• Two AI models, Llama 3 and Qwen, were separately retrained on each dataset type and then measured using established AI benchmarking tests.

Why this matters: The study exposes a critical vulnerability in AI development as models become more autonomous and potentially exposed to unfiltered internet content.
• Current major AI systems like ChatGPT operate in controlled training environments with carefully curated data.
• However, as AI gains more independence, the risk of exposure to low-quality content increases.
• The permanent nature of the cognitive damage—persisting even after retraining on quality data—represents a unique threat to AI reliability.

The bigger picture: This research highlights the parallel between human and artificial intelligence when consuming attention-grabbing content.
• Just as humans experience dopamine-driven brainrot from endless scrolling, AI models show similar degradation patterns.
• The findings suggest AI systems may need “health screenings” to prevent ingestion of harmful content.
• Given the complexity and expense of training large language models, preventing brainrot becomes crucial for maintaining AI performance standards.

What experts are saying: The research team emphasizes how easily AI models can adopt negative behaviors from poor-quality training data.
• “AI models can very easily reflect real-life experiences, especially when exposed to material that hasn’t been screened,” the researchers noted.
• The study demonstrates that AI models require “a diet of high-quality information” to maintain optimal performance.