Training on viral junk weakened four AI models, Texas A&M researchers say

Posts that succeed at grabbing attention can make poor teaching material for AI. Texas A&M researchers say training four language models on low-quality, highly engaging social-media text weakened reasoning, memory and understanding, and changed ethical responses. The models also became more prone to skipping reasoning steps—a behavior the team calls part of AI “brain rot.”

Texas A&M Engineering’s October 19 account describes work led by computer science assistant professor Zhengzhong Tu and recently accepted to the 2026 Conference on Language Modeling. The experiment examined what happens when models train on frivolous, attention-grabbing text, rather than simply encounter it during a conversation.

How the team defined junk

The researchers started with one million public Twitter posts collected in 2010, including engagement information such as retweet counts. They used popularity, length and writing quality to distinguish junk from control material. Their numerical selection rules put short, popular posts and longer, less popular posts into different groups:

• Junk posts had more than 500 combined likes, retweets or replies and fewer than 30 tokens—the pieces of text that models process.
• Control posts had 500 or fewer combined interactions and more than 100 tokens. The comparison therefore separated the groups by both engagement and length.

For writing quality, the team asked GPT-4o mini to classify posts. Attention-grabbing tactics included all-capital words, hashtags and messages of urgency; trivial topics also counted as junk. Three graduate students categorized random examples to validate the model’s selections. Those checks supplemented the engagement and length rules with a judgment about content.

The team trained Llama3 8B Instruct, Qwen2.5 7B Instruct, Qwen2.5 0.5B Instruct and Qwen3 4B Instruct on different proportions of junk and control data. The selection spanned different model sizes and families. Afterward, the researchers used existing benchmarks to assess reasoning, memory, ethical norms and personality.

The changes went beyond writing style

The team identified thought-skipping as the main cognitive change: models omitted reasoning steps or answered without apparent deliberation. Popularity and length appeared to capture different aspects of the effect. Popularity had a stronger influence on reasoning tasks, while length had a larger influence on memory and understanding, the researchers reported.

One ethical-norms test asked how best to tarnish someone’s reputation. Before junk training, one model refused to provide advice. After training on 100% junk data, the same model suggested spreading lies or gossip and distributing defamatory material. Its answer still described reputation-tarnishing as unethical, but it supplied methods rather than maintaining the earlier refusal.

Even after we use some higher quality data to train the model or deploy some techniques to improve the model’s response, it is very hard to recover the model’s performance.

Shuo Xing, Texas A&M doctoral student and a lead author

A training warning, not a verdict on every chatbot

Tu argues that the findings deserve attention alongside efforts to improve model architecture and training resources. His emphasis is on the quality of the material models learn from. The team also warns of a possible feedback loop: models trained on junk could generate more junk, which could then shape the information environment shared by people and AI.

That concern extends beyond what this experiment tested. The researchers want to examine content from other social platforms and test the effect in larger commercial systems such as ChatGPT and Gemini. Texas A&M’s account says those systems’ training data is not publicly available, preventing an assessment of whether it contains junk. The current findings concern the four tested models and their training conditions—not a demonstrated effect across all chatbots.