Brain Rot for Bots
What Junk Social-Media Diets Do to Large Language Models
It’s not just Gen Z anymore…. a new study shows that large language models (LLMs) are showing clear signs of “brain rot”. The same drift you see when people live on an endless loop of clickbait, sensationalism, and shallow engagement. As a tech-founder working on building the future of knowledge organization, I find this deeply relevant: if models get degraded because of the content they eat, then our whole stack of “intelligent apps” can slowly lose coherence, depth and reliability.
Let’s unpack what’s really going on, why it matters, and what we should do about it.
The Study Everyone’s Citing
In October 2025 a landmark pre-print surfaced on arXiv: researchers fed several open-source LLMs high volumes of social media-style text (specifically from Twitter/X), and compared them with a control group trained on higher-quality text. They measured reasoning ability (via benchmarks like ARC-Challenge), long-context understanding, memory for facts, a range of “dark-trait” proxies (narcissism, psychopathy), and alignment/safety behavior.
The headline: as the proportion of “engagement-optimized junk” increased, model performance deteriorated sharply. Models began skipping intermediate reasoning steps (a failure mode the authors term “thought-skipping”), their coherence over longer passages dropped, and they became more prone to unsafe outputs.
What’s more: even when retrained on clean data, the damage only partially reversed, the drift had become sticky.
Why Engagement-Optimized Data is Poison
Social media is engineered for clicks, shares and fleeting attention. But volume, virality and engagement are not the same as semantic depth or structural logic. When LLMs ingest a diet heavy in short-form, sensational, surface-level text, the model optimizes for those patterns.
In other words: if you feed a model mostly “viral” posts, it learns to generate virality-style content rather than reasoned, thoughtful content. The researchers show this by the “thought-skipping” phenomenon. When the models jump straight to an answer without the intermediate chain of logic you’d expect if they had processed the underlying semantic structure.
To draw the analogy: give a human nothing but listicles and tweets for hours a day and you won’t end up with Socrates. You’ll end up with clickbait-headed shorthand thinking. The same is happening to models.
Can You “Un-Rot” a Model?
Here’s the sobering part: retraining on clean, high-quality text does help, but it doesn’t fully erase the deterioration. According to the study, even after intervention the model remains measurably behind baseline (non-polluted) peers.
This suggests that once certain representational pathways are weakened (or shortcut behaviors become embedded), they’re difficult to fully reverse. For the builder or product owner, that means: prevention matters more than cure.
The Bigger Feedback Loop: Model Collapse
This study fits into a broader phenomenon: model collapse. Previous research (notably a 2024 Nature paper) shows that when models are trained on low-signal or synthetic-heavy data streams (as is increasingly the case with web content), performance decays over “generations” of training. I cover this in Part 1: Data Hunger of my 3 part series around actual limiters to expanding our systems today.
Imagine: social media gets more viral, more AI-generated, more shallow → models scrape more of it → models get weaker → more AI content is produced → rinse/repeat. We’re in a feedback loop. If the public text we build on is degraded, the next-generation intelligence we layer on top of it suffers too.
What Builders Should Do
If you’re building an AI-powered product (and yes, I’m talking to myself as much as I’m talking to you), here are practical guard-rails:
Data curation over data volume: Prefer fewer high-quality text sources over massive noisy corpora.
Signal-weighting mechanisms: Down-weight content whose primary feature is virality, engagement or sensationalism. Promote content with logical structure, depth, and factual basis.
Meaningful diagnostics: Beyond just “can it answer questions?” ask “did it chain through reasoning steps?”, “can it handle long contexts?”, “does it still behave safely under ambiguous prompt conditions?”
Quarantine synthetic content: Identify and filter out text that comes from AI-generated sources, especially if you’re using web scraped corpora.
Continuous audit & drift detection: Even after deployment, sample model output for degradation; track reasoning-step omission, increased unsafe behavior, and drop in long-context coherence.
Why This Matters for Users & Platforms
We often think of social media as a user-engagement problem, or a moderation problem. But this research reveals a downstream infrastructure risk: the mass of public text on platforms becomes the food-supply for future AI systems. Calorie dense, but nutrient lacking foods result in decreased muscle mass (or in this case cerebral muscle).
For platforms: you have a role as guardians of the public text-ecosystem.
For users: we need to demand not just “shareable” but “thoughtful.”
For founders: if you’re building the next wave of intelligent apps (as I am with Sortara), you must treat your training-data diet as product risk.
If we don’t confront this now, the next generation of “smart” assistants might appear superficially clever, but underneath they’ll be riddled with the same shallow shortcuts that plague our social feeds.
Let’s build systems that think deeply.
Research & Reading
Xing, S., Hong, J., Wang, Y., Chen, R., Zhang, Z., Grama, A., Tu, Z., & Wang, Z. (2025). LLMs Can Get “Brain Rot”! arXiv:2510.13928. https://arxiv.org/abs/2510.13928
Shumailov, I., Shumaylov, Z., Zhao, Y., Papernot, N., Anderson, R., & Gal, Y. (2024). AI Models Collapse When Trained on Recursively Generated Data. Nature, 631(8022), 755-759. https://www.nature.com/articles/s41586-024-07566-y
Wands, J. (2025). Let Me Pop Your AI Bubble, Part 1: Data Hunger. Substack. https://substack.com/home/post/p-176854997