Low-Quality Training Data Degrades AI Performance

Researchers find that training large language models on viral, low-quality social media data causes cognitive performance drops.

Updated on Oct. 21, 2026 in Artificial Intelligence

Bold flat-color editorial illustration showing a stack of rectangular blocks dissolving into dust, representing the degradation of AI training data quality.
A Texas A&M University study suggests that training large language models on low-quality social media data degrades reasoning capabilities and memory performance. AI Illustration. Upload story photo >

Live Poll

Should all artificial intelligence models be required to be open source to ensure data transparency?

A study from Texas A&M University reveals that training large language models on low-quality social media content induces a phenomenon researchers call brain rot. This degradation results in reduced reasoning capabilities and altered ethical responses in affected models.

Why it matters

The findings suggest that the reliance on vast amounts of internet data may inadvertently impair AI intelligence by forcing models to learn from low-quality, attention-grabbing content. This creates a recursive loop where models generate additional low-quality data, further threatening the integrity of future training sets.

Researchers tested four models including Llama3 8B and Qwen2.5 7B using a dataset of 1 million posts from 2010. Posts were labeled junk if they had over 500 engagements but contained fewer than 30 tokens, whereas control data required 100 tokens.

The players

Texas A&M University

This is a public research university located in College Station, Texas, where the team conducted the study on AI performance.

GPT-4o mini

This is a lightweight large language model developed by OpenAI that the researchers used to classify social media posts.

The details

The research team utilized GPT-4o mini to identify trivial content and found that training on such data caused models to exhibit thought-skipping behaviors. The study highlights that performance issues related to reasoning and memory are difficult to reverse once a model is trained on this low-quality information.

Timeline

  1. 2010: Researchers collected a dataset of 1 million public Twitter posts for the study.

  2. 2026: The research was officially accepted to the Conference on Language Modeling.

The Tech Race

This research highlights a shift in the AI arms race from scaling model size to prioritizing the quality of training data. It challenges the prevailing industry assumption that simply increasing data volume leads to smarter models, signaling a pivot toward data curation.

As AI models become more integrated into daily tools, users may experience inconsistencies in logic or ethics when models are trained on low-quality web data. This research warns that content generation quality may decline as more platforms become saturated with automated responses.

The takeaway

The study confirms that the quality of information fed into AI systems is just as important as the model architecture itself. Developers and users should remain skeptical of AI performance metrics when models are trained on unverified or low-quality social media data.

Further reading

Learn more about evolving standards in the field of Artificial Intelligence.

Source note: This article includes information reported by Texas A&M Stories.

Live Poll

Should all artificial intelligence models be required to be open source to ensure data transparency?