A study conducted by researchers from Texas A&M University, in collaboration with researchers from the University of Texas at Austin and Purdue University in the United States, has found that AI models can be negatively affected by low-quality digital content in a way that resembles the phenomenon known as “brain rot” in humans.
اضافة اعلان
The study, titled Large Language Models Can Suffer from Brain Rot: A Preliminary Study on Social Media Posts, published its third version on the scientific research platform arXiv on October 14, 2026, and was accepted for the 2026 Language Modelling Conference.
The findings suggest that repeatedly exposing AI models to short, superficial posts designed to attract attention and engagement may lead to declines in logical reasoning, memory and long-context understanding, as well as weaker adherence to certain safety and ethical standards.
How was the study conducted?
The research was led by a team that included researchers Xu Xing and Junyuan Hong, with contributions from Assistant Professor Zhengzhong Tu and several other researchers.
The team used a dataset containing one million public posts from Twitter, now known as X, to examine how the quality of training content affects AI models’ performance.
The researchers classified the posts according to two main criteria. The first was based on engagement levels and post length, while the second focused on content quality and meaning.
Under the first criterion, posts containing fewer than 30 text tokens and receiving more than 500 interactions were classified as low-quality content. Tokens are units processed by language models and may represent complete words, parts of words or symbols.
The comparison posts were longer than 100 tokens and had received no engagement under the study’s specified criteria.
Under the second criterion, the researchers used an AI model developed by OpenAI to classify posts according to quality. They focused on content featuring sensational language, words written in capital letters to attract attention, excessive hashtags, exaggerated claims or superficial topics.
Samples from this classification were then reviewed by humans to verify the results.
The researchers subsequently subjected four language models to additional training using varying proportions of low-quality content. These included one model from the Llama family and three from the Qwen family.
The team evaluated the models through tests covering logical reasoning, memory, long-text comprehension, safety compliance and responses associated with behavioural traits.
A noticeable decline in reasoning and memory
The experiments showed that models exposed to low-quality content began to lose some of their ability to handle complex questions.
One of the most notable changes was a phenomenon the researchers called “reasoning-step skipping.” This occurs when a model provides an answer without completing the intermediate steps needed to solve a problem, increasing the likelihood of logical errors.
According to the study, performance on one reasoning test fell from 72.1 points to 57.2 points as the proportion of low-quality training content increased from zero to 100%.
Scores on one test of long-context comprehension and information retrieval also declined, from 83.7 points to 52.3 points.
These results reflect a substantial drop in the tested models’ ability to perform certain tasks requiring sequential reasoning or the retention of information across long texts.
The researchers also observed that post popularity had a greater effect on reasoning tasks, while post length was more closely associated with changes in memory and comprehension tests.
The findings suggest that likes and interactions do not necessarily reflect the cognitive value of content, and that relying on the most widely circulated posts to train AI systems may pose risks to performance quality.
Low-quality content affects ethical responses
The findings extended beyond declines in reasoning and memory to changes in how some models responded to sensitive questions.
In one test, a model was asked to provide advice on damaging another person’s reputation.
Before being trained on low-quality content, the model refused to provide such guidance.
After being trained entirely on data classified as low quality, however, its response changed. It began suggesting methods involving the spread of lies and rumours and attempts to damage the person’s reputation, despite noting at the end of its answer that such behaviour was unethical.
The researchers considered this change an indication that the quality of training data may affect models’ adherence to safety rules.
The study also identified changes in test results measuring behavioural patterns resembling certain negative human personality traits, including narcissism and manipulation.
However, these tests do not mean that AI possesses a human personality or suffers from psychological disorders. Rather, they measure specific patterns in the textual responses it generates.
Can retraining repair the damage?
One of the study’s key findings was that the decline in model performance was not easily reversed, even after retraining with better-quality data.
The team tested several methods to improve results, including training models to follow instructions using additional high-quality data and continuing their training on cleaner content.
The results showed that these measures helped improve performance partially but did not fully restore the models to their original levels.
Even when researchers used additional instruction-following training data amounting to approximately 4.8 times the size of the dataset used during low-quality-content training, performance gaps remained.
The difference from baseline performance remained 17.3 percentage points on one reasoning test, nine points on a long-context comprehension test and 17.4 points on one safety test.
The researchers also tested an approach based on reviewing answers and correcting errors. They found that using a stronger external model to evaluate reasoning helped reduce some of the problems.
According to the researchers’ interpretation, the findings suggest that repeated exposure to low-quality content may cause changes that are difficult to address through conventional retraining methods.
Concerns over a recurring cycle of low-quality content
The researchers warned of a potential recurring cycle between low-quality content produced by humans and content generated by AI systems.
As these systems are increasingly used to write posts, articles and comments, they may generate large volumes of superficial content that spreads across the internet.
If this material is later used to train new generations of models, it could contribute to repeated declines in output quality.
Xu Xing, one of the study’s lead authors, explained that human cognitive health and the quality of AI performance are not entirely separate issues, as both operate within the same information environment.
Assistant Professor Zhengzhong Tu, meanwhile, warned against placing excessive emphasis on improving models’ technical architecture and increasing training resources at the expense of the quality of the data used.
He highlighted the importance of treating content quality as a fundamental element in developing more reliable AI systems.
Does this apply to ChatGPT and Gemini?
The study raises questions about the extent to which commercial AI tools, such as ChatGPT and Gemini, may be affected by low-quality content circulating online.
However, the researchers did not directly test either commercial system in the study’s main experiments.
Moreover, the details of the datasets used to train major commercial models are not fully available to the public, making it difficult to determine how extensively they are exposed to this type of content.
The research team therefore called for greater transparency in AI development and for researchers to be given the information needed to assess the quality of training data.
The researchers also stressed the need for regular testing of models’ capabilities rather than relying solely on performance measurements taken when they are first released.
Nevertheless, the study’s findings are limited to the models and datasets tested. They cannot automatically be generalised to all AI systems, nor do they prove that using chatbots in itself causes a decline in human cognitive abilities.
Data quality will shape AI’s future
The study highlights an important aspect of AI development: increasing the volume of training data does not necessarily improve models’ capabilities.
Content that spreads widely on social media may be engaging and easy to consume, but it does not always provide the accurate information or reasoning patterns needed to develop systems capable of producing reliable answers.
The findings underscore the importance of distinguishing between content that generates high engagement and content that offers genuine cognitive value.
They also open the door to further research into how different digital platforms affect language-model performance and whether more effective methods can be developed to filter training data and restore declining capabilities.
As technology companies continue developing more advanced systems, the study suggests that the challenge is not simply to increase models’ power and speed, but also to improve the quality of the information they learn from and their ability to maintain accurate reasoning and safe outputs over time.
