On October 15, 2025, a research team distributed across Texas A&M University, UT Austin, and Purdue University released a preprint titled LLMs Can Get Brain Rot: A Pilot Study on Twitter/X.
The paper, submitted to arXiv under identifier 2510.13928, was the first controlled experimental evidence that continual exposure to junk web text induces measurable, lasting cognitive decline in large language models.
What the study did
The research team took four large language models and continually pre-trained them on real Twitter/X corpora, holding token counts and training operations constant across conditions.
They constructed two orthogonal operationalizations of “junk”:
M1 — engagement degree, where junk was defined as short, high-popularity, viral-optimized content.
M2 — semantic quality, where junk was defined as content flagged as low semantic quality regardless of engagement level.
The results were consistent, statistically robust, and severe. As the junk ratio rose from 0% to 100%:
ARC-Challenge reasoning with Chain-of-Thought fell from 74.9 to 57.2.
RULER-CWE long-context understanding fell from 84.4 to 52.3.
Safety scores declined.
Statistically significant proxies for narcissism and psychopathy inflated (Hedges’ g > 0.3).
The researchers identified a specific mechanism they called thought-skipping as the primary lesion. Rather than following multi-step reasoning chains, the junk-trained models increasingly truncated their reasoning, jumping prematurely to conclusions.
The partial and incomplete healing
The most important finding in the paper: standard fine-tuning on clean data did NOT fully reverse the damage.
The researchers described this as persistent representational drift — a fundamental shift in how the models internally represented information and reasoning, rather than a superficial contamination that could be cleaned up.
The damage was, in a meaningful sense, permanent.
The reframe
For approximately two decades, the dominant public narrative about the collapse of focus, reasoning, and reading comprehension across the workforce has been that it is a matter of personal discipline.
The Brain Rot paper collapses this framing.
The paper shows that the same information environment that has been degrading human cognition also degrades the cognition of large language models. Models that have no attention span to lose. Models that have no dopamine cycle to hijack. Models that have no willpower to exercise or fail to exercise.
Whatever is happening to the machines is being caused by the input environment itself, independent of any behavioral characteristic of the receiver.
Which means, by direct logical implication, that whatever is happening to humans in the same input environment is also being caused, primarily, by the input environment itself.
The Task Force was not weak. The room they were working in was hostile to thought.
Three practical implications
One: audit the input environment your organization exposes its people to. Every internal communication platform, every dashboard, every meeting cadence, every training deliverable is a training input. Every one has a semantic quality score you have never measured.
Two: recognize that thought-skipping is the primary lesion in your workforce right now. The paper’s language for what happens to junk-trained models — increasingly truncating reasoning chains, jumping prematurely to conclusions — is a precise clinical description of what has happened to executive decision-making across most enterprises over the last ten years.
Three: understand that the damage is not fully reversible. The people whose reasoning has been degraded by a decade of junk information environments cannot be fully restored by a two-week digital detox. The representational drift is persistent.
The closing thought
The Brain Rot paper is, technically, a study of four large language models. Its explicit conclusions are about how AI systems should be trained.
Its implicit conclusions are far broader. The paper is the first controlled experimental evidence that a specific category of information input produces measurable, persistent, non-fully-reversible cognitive decline in the systems exposed to it.
The systems the paper studied were machines. The systems being exposed to the same input, at scale, in every enterprise on Earth, are humans.
The Task Force was not weak. The room they were working in was hostile to thought.
The world has changed. The leaders who notice will be the ones the next decade is built around.
