Researchers Warn of AI Deterioration from Poor Social Media Data

Researchers from UT Austin, Texas A&M, and Purdue found large language models can develop AI brain rot from junk social media text.
Models like Llama and Qwen were trained on HuggingFace's tweets and clickbait.
Benchmarks showed increased thought-skipping and lower ethical scores.
Llama3 and Qwen had reduced accuracy after junk exposure.
Researchers recommend prioritizing data quality and routine evaluations.
AI-generated content risks worsening 'brain rot', linked to Oxford's 2024 Word of the Year.

AI Ethics | Social Media | Language Models

United States | United Kingdom

Junyuan Hong

2 months ago