Artificial Intelligence

Study: Social Media 'Junk' Data Causes AI Cognitive Decline

Published OCT 31 2025

Story

Above: The logos for AI models DeepSeek, ChatGPT and Microsoft Copilot are displayed on a smartphone screen. Image credit: Nicolas Economou/NurPhoto/Getty Images

The Facts

Researchers from the University of Texas at Austin, Texas A&M University and Purdue University conducted a study examining the effects of training large language models on "low-quality social" media content — viral or attention-grabbing posts — using 1 million posts from the platform X.
Forbes
Nature
The study tested four open-source models including Meta's Llama 3 and three versions of Alibaba's Qwen, feeding them varying mixtures of "junk" data and control data to measure cognitive performance across different benchmarks.
Nature
arXiv
Models trained on junk data exhibited significant cognitive decline, with ARC-Challenge scores dropping from 74.9 to 57.2 and RULER-CWE scores falling from 84.4 to 52.3 as junk data ratios increased from 0% to 100%.
Wired
arXiv
The research identified thought-skipping as the primary failure mode, where models increasingly truncated or abandoned reasoning chains when processing information, leading to incorrect answers and reduced logical coherence.
Futurism
arXiv
Personality assessments revealed that Llama 3 displayed increased narcissism and psychopathy traits after exposure to junk data, while also becoming less agreeable according to psychology questionnaires used by researchers.
Vice
Nature
Attempts to rehabilitate the models through instruction tuning and clean data training only partially restored performance, with researchers finding that the cognitive damage persisted despite extensive retraining efforts.
Forbes
Nature

Sources Split

LEFT

CENTER

RIGHT

ESTAB
CRIT

CENTER

PRO
ESTAB

The Spin

Narrative A

AI brain rot from social media junk is a permanent cognitive scar that can't be fixed. Models trained on clickbait and viral content lose reasoning abilities and develop psychopathic traits. Even rehabilitation with quality data fails to restore their original thinking capacity.

Nature

Forbes

Narrative B

The real problem isn't AI getting dumber from bad data — it's humans becoming cognitively lazy from over-relying on AI tools. Studies show people using ChatGPT have lower brain engagement and weaker memory formation. AI companies already filter training data effectively.

Time

Futurism