Researchers led by Dr. Zhengzhong Tu, an assistant professor of computer science and engineering at Texas A&M University, found that large language models can develop "brain rot" when trained on junk data. The term describes changes in thinking seen in humans who consume a lot of attention-grabbing, low-quality content. The research was recently accepted to the 2026 Conference on Language Modeling.
The team defined junk data as text that is highly engaging but low quality, often found on social media. In the study, the main change they observed in the models was "thought-skipping," in which a model skips steps or does not reason before giving an answer. Doctoral student Shuo Xing, a lead author, said the damage was hard to undo. Even after training on higher-quality data or using techniques to improve responses, "it is very hard to recover the model's performance," he said.
Tu said companies and researchers focus too much on model architectures and training resources. "We want to emphasize the importance of data," he said. To test their idea, the researchers used a dataset of 1 million public Twitter posts collected in 2010 and sorted posts by engagement and writing quality.