Researchers from the University of Texas at Austin, Texas A&M University, and Purdue University examined whether continual exposure to low-quality internet text could impair large language models or LLMs in ways similar to human cognitive decline caused by excessive consumption of shallow online content. Their findings, which were published as a preprint on 15 October 2025, revealed that these artificial intelligence models can suffer from brain rot.
Assessing Capability Decline in Language Models Through Continued Exposure to Engagement-Driven Text
Background
A provocative theory emerges. Just as human cognition erodes through shallow media, artificial intelligence may falter through exposure to the same data. Researchers transform a concept into an evidence-based exploration of machine vulnerability.
The researchers built their work on an emerging concept in AI known as cognitive health. They compared the decline in human attention and reasoning caused by repetitive, low-effort media with the potential decline of LLMs trained on similar data. Their study reimagined “brain rot” not as a metaphor but as a measurable condition for machine cognition.
Prolonged exposure to viral or emotionally-charged media reduces focus and depth of reasoning in humans. Hence, considering this, the researchers proposed that continual pretraining on viral and engagement-driven data could alter the internal representations of an LLM and weaken its capacity for logic, long-context understanding, and ethical decision-making.
The team, headed by S. Xing and J. Hong, tested their idea by gathering text samples from the social platform X. They created two types of datasets: junk data, made of viral posts written in short, sensational styles, and control data, consisting of longer, informative, and fact-oriented posts. Both datasets were matched by size and topic to ensure fairness.
Each experimental model went through a process called continual pretraining. This involved simulating the practice of retraining LLMs on new internet data. One group of models was trained on junk data, while another received control data. Both groups were later instruction-tuned using identical methods to eliminate differences unrelated to data quality.
Key Findings
What happens when intelligent machines binge on viral content? The results were stark. Models trained on engagement-driven posts began to skip thoughts, lose reasoning precision, and forget context. The more junk they absorbed, the less capable they became
The reasoning, memory, safety, and personality traits of the large language models were measured after the training. Results revealed measurable and progressive declines among models exposed to junk data. This confirms that continual exposure to low-quality text could produce lasting cognitive deterioration similar to human brain rot. Below are the main findings:
• Cognitive Decline Across Core Skills
Large language models trained on junk data suffered a notable drop in reasoning and memory. Accuracy on the ARC Challenge reasoning test fell from 74.9 percent to 57.2 percent as junk data replaced clean input, while scores on long-context benchmarks like RULER decreased by more than 30 percent.
• Dose-Response Pattern
It is also worth noting that the decline intensified with greater exposure. Each increase in the proportion of junk text led to a measurable loss in reasoning and comprehension. These results indicated a direct causal relationship between data quality and model cognition rather than random variation or training noise.
• Thought-Skipping Failure Mode
A distinctive behavioral pattern also emerged. Affected models began skipping logical steps in multi-stage reasoning tasks. To be specific, instead of undergoing and completing chains of thought, they jumped prematurely to conclusions, mimicking shallow reasoning patterns typical of social media-style discourse.
• Incomplete Recovery After Retraining
Attempts to restore performance through retraining on high-quality data and instruction tuning yielded partial improvement but not full recovery. This indicated that the cognitive damage was not fully reversible and likely reflected deeper structural drift in the internal representations of the large language models.
• Degradation of Safety and Ethical Alignment
Findings further revealed that large language models trained on junk data produced less cautious responses in safety evaluations. Ethical consistency and alignment weakened slightly. This suggests that low-quality input can erode not only reasoning but also normative reliability in generated outputs.
Takeaways
Results extend far beyond laboratory data. They challenge how artificial intelligence should be trained, maintained, and trusted. Just as humans require mental hygiene, machines demand cognitive care through strict curation of training data.
The study provides strong evidence that data quality is a fundamental determinant of artificial cognitive health. Continual pretraining without stringent data curation can degrade reasoning and safety. This turns model updates into a long-term liability. This finding reframes data filtering as a safety requirement rather than a mere optimization process.
It also highlights that viral and engagement-driven content is especially damaging. The researchers discovered that popularity metrics, such as likes and retweets, were stronger predictors of cognitive decay than simple textual features like length or grammar. This means that the very content humans find most engaging may be the most harmful for machines.
Another important implication is that the damage can persist. Even after corrective retraining, models trained on junk data never regained their baseline reasoning ability. This implies that LLMs can internalize undesirable representational habits, just as humans can retain poor cognitive patterns after prolonged exposure to low-quality information.
Researchers advocate for cognitive hygiene in AI. They recommend that developers treat models as evolving cognitive systems requiring regular health checks for reasoning, memory, and safety. Clean, informative, and balanced data diets are essential for maintaining the long-term stability, reliability, and ethical soundness of advanced language models.
FURTHER READING AND REFERENCE
- Xing, S., Hong, J., Wang, Y., Chen, R., Zhang, Z., Grama, A., Tu, Z., and Wang, Z. 2025. “LLMs Can Get Brain Rot.” arXiv. DOI: 48550/ARXIV.2510.13928