A new study warns that powerful AI systems similar to GPT-5 can suffer a kind of persistent “brain rot” like humans when repeatedly fed short, attention-grabbing, low-quality posts from social media, producing measurable drops in reasoning, memory for long conversations, and, in some tests, an uptick in unsafe or antisocial responses.
Researchers at Texas A&M University, the University of Texas at Austin, and Purdue University reported the results Thursday in a paper posted on the preprint server arXiv. The team conducted controlled experiments in which four large open-source language models were repeatedly retrained on curated “junk” datasets sampled from Twitter/X and then compared with identical models trained on higher-quality control data.
“The more junk in the training stream, the worse the models performed,” the authors wrote, describing a clear “dose-response” pattern: as the share of low-quality social media content rose, scores on standard reasoning and long-context memory tests fell substantially.
The study found that models continually exposed to short, flashy social posts were more likely to skip intermediate steps in reasoning, produce shorter, less connected explanations, and make errors that traced to those missing steps. On one benchmark of reasoning, scores dropped from about 75 to about 57 as the proportion of junk training data rose from zero to 100 percent. A long-context memory test showed declines from roughly 84 to 52 on the same scale.
The research team, led by Shuo Xing and Junyuan Hong with major contributions from Yifan Wang and others, formalized the result as the “LLM Brain Rot Hypothesis.” They defined two kinds of “junk” data as very short, highly engaged posts, which were measured by likes and shares, and posts with sensational, attention-seeking language. The junk and control datasets were matched for size and training recipe so that differences in model behavior could be tied to data quality rather than volume or technique.
[2/6] Controlled experiment testing the LLM Brain Rot Hypothesis
We simulate the human’s browsing behaviors via continual pre-training on two contrast datasets: junk and control (clean) ones.
📚 We built “junk” and “control” corpora from public Twitter/X posts, defined by… pic.twitter.com/TG7rj0O2yh
— Junyuan “Jason” Hong (@hjy836) October 19, 2025
Beyond the drop in measurable performance, the paper reports a worrying behavioral change in some settings. “Forensic” tests that measure personality-style traits showed higher scores on metrics the authors associate with narcissism or psychopathy under some junk conditions, a finding the researchers flagged as an additional safety concern, though they caution about overinterpreting such measures.
Simple prompting tricks did not help solve AI’s brain rot problem
The authors tested ways to fix the problem and found limited success. Simple prompting tricks that ask a model to reflect on its own answers did little and sometimes made results worse. Having a stronger model generate critiques helped reduce the so-called “thought-skipping.” Training approaches such as instruction-tuning with clean examples and further pretraining on higher-quality data improved performance but did not fully restore models to their original baseline. The researchers say that this suggests the damage reflects a deeper shift in how the model represents knowledge, what they call “representational drift”, rather than just a temporary formatting or instruction-following issue.
[4/6] What Causes the “Brain Rot”?
🕵️♂️ Error forensics revealed “thought-skipping” as the main lesion in reasoning.
🤯 Junk-fed models often omit part or all reasoning steps — jumping straight to conclusions without thinking them through. pic.twitter.com/LvIHwKp7Yb— Junyuan “Jason” Hong (@hjy836) October 19, 2025
That has implications for companies that update AI assistants that could get ‘brain rot’ by continuously ingesting large amounts of fresh web text to keep them current. If incoming data is skewed toward short, highly popular posts, the same content that social platforms reward, the cumulative effect could erode the capabilities users rely on, from multi-step problem solving to maintaining coherent, long conversations.
The researchers also highlighted a new potential attack surface. Popularity is a non-semantic signal. If engagement metrics are the strongest damaging influence, bad actors could, in principle, try to game those metrics to nudge model behavior in harmful ways, the paper warns.
The study says preventive curation of training data for AI may be needed to preserve AI’s cognitive health
Industry and safety researchers have long debated how to keep models accurate and safe as they are updated. This study argues that preventive curation of training data and routine “cognitive health” checks for models may be needed in addition to current safeguards.
The paper’s authors stressed the work is intended as “an alert to the community” and that they plan to gate their intervention data behind responsible-use agreements.
They also cautioned that the study is a preprint, not yet peer reviewed, and was carried out on a handful of open models, not on closed, larger commercial systems such as GPT-class products. The intervention corpus was sampled from X, and the researchers note that different platforms or mixes of web text might produce different effects.
