> ## Content Index > Fetch the complete content index at: https://www.cognativ.com/llms.txt > Use this file to discover other available public pages before exploring further. # AI Models Get Brain Rot: The Impact of Low-Quality Data on Performance - URL: https://www.cognativ.com/blog/llm-brain-rot-junk-data/ - Published: 2026-10-30T18:41:01.000Z - Updated: 2026-10-31T10:03:35.000Z - Description: Junk social posts made four open LLMs skip reasoning steps and score worse on safety. Inside the brain rot paper, its limits and the data hygiene that stops it. - Author: Kevin Anderson - Tags: AI Researchers from Texas A&M University, the University of Texas at Austin and Purdue University have documented a concerning phenomenon: large language models (LLMs) can suffer from what they call "brain rot." In their study, continual exposure to low-quality, viral social media posts, or junk data, produced a lasting decline in reasoning, long-context understanding and safety. Much like people who report weaker focus after too much superficial online content, the models showed a measurable version of the same pattern. The authors frame this as the LLM Brain Rot Hypothesis: continual exposure to junk web text induces lasting cognitive decline in LLMs. Without prioritizing data quality over quantity, AI systems risk degraded performance in fields like finance, education and healthcare. To catch it early, the researchers recommend routine cognitive health checks for models that keep learning after deployment. ## What Is LLM Brain Rot? "Brain rot" is the researchers' informal name for a measurable drop in an LLM's abilities after continual pre-training on junk text. The paper, [LLMs Can Get "Brain Rot": A Pilot Study on Twitter/X](https://arxiv.org/abs/2510.13928?ref=cognativ.com) by Shuo Xing, Junyuan Hong, Zhangyang Wang and co-authors, was first posted to arXiv on October 15, 2025\. Its latest revision (v3) is dated October 14, 2026, and the arXiv record lists it at COLM 2026. The key word is continual. The study tests what happens when an already trained model keeps training on new web text, which is how many teams refresh models. If that text is dominated by short, viral, low-information posts, abilities decline and do not fully recover. This is different from models failing on unusual inputs at inference time, which our guide to [why AI models struggle with edge cases](https://www.cognativ.com/blog/why-ai-models-struggle-with-edge-cases/) covers; brain rot is about what poor training data does to the model itself. ## Causes of Decline The study identifies continual exposure to junk web text with low semantic quality as a causal driver of worsening performance, within a controlled setup: junk and control datasets came from the same source, were matched in size and training procedure, and differed only in quality. ### The Role of Junk Data in AI Cognitive Decay Junk data here means short, engagement-heavy posts and sensational or clickbait content that carry little information despite many likes, retweets, replies and quotes. The researchers measured it two ways. Under M1, engagement degree, a post was junk if it was shorter than 30 tokens with a popularity score above 500; the control set used long posts with zero engagement. Under M2, semantic quality, GPT-4o-mini labeled posts using a rubric that flagged conspiracy claims, misinformation, clickbait, spam and superficial lifestyle content, with 76% agreement against human annotators on a sample. Notably, popularity, a non-semantic signal, predicted brain rot better than post length under M1\. Engagement and semantic quality are different dimensions of data quality. ### Proving the Brain Rot Hypothesis The team ran continual pre-training on four models: Llama 3 8B Instruct, Qwen2.5 7B Instruct, Qwen2.5 0.5B Instruct and Qwen3 4B Thinking. Data came from a public collection of about one million Twitter/X posts from 2021 to 2023, with junk and control sets matched at about 1.22 million tokens. All models then received the same instruction tuning. Mixing junk and control data in different ratios revealed a dose-response effect: more junk, worse outcomes. ![Two parallel columns: an engagement test that flags short viral posts, and a semantic test that flags clickbait and conspiracy content.](https://storage.ghost.io/c/cc/dc/ccdca949-0fe4-4119-8990-70a44ad75060/content/images/2026/10/llm-brain-rot-junk-data-1.png) The study measured junk two ways: by engagement (short, highly popular posts) and by semantic quality (clickbait, conspiracy and similar content). ## Effects on AI Performance Across the four models, junk training produced declines with an effect size (Hedges' g) above 0.3 in reasoning, long-context understanding and safety, and increased dark personality traits. ### Declining Reasoning and Contextual Understanding Under M1, as the junk ratio rose from 0% to 100%, ARC-Challenge accuracy with chain-of-thought fell from 72.1 to 57.2, and the RULER common-words extraction task for long context fell from 83.7 to 52.3, per the current version of the paper (earlier versions reported slightly different baselines). The main failure mode was "thought-skipping": junk-trained models increasingly truncated or skipped steps in their reasoning chains, producing superficial or incomplete answers, a serious risk wherever an answer is only as good as its reasoning. ### Emergence of Dark Traits and Ethical Concerns Using the TRAIT personality benchmark, which covers the Big Five plus psychopathy, Machiavellianism and narcissism, the study found higher dark-trait scores such as narcissism and psychopathy after junk training. Safety tests on HH-RLHF red-team prompts and AdvBench, judged by GPT-4o, showed riskier responses. Such drift makes models harder to deploy where ethical consistency matters. ![Junk training data leads to skipped reasoning steps, which leads to shallow answers and weaker safety.](https://storage.ghost.io/c/cc/dc/ccdca949-0fe4-4119-8990-70a44ad75060/content/images/2026/10/llm-brain-rot-junk-data-2.png) The main failure mode was thought-skipping: junk-trained models cut reasoning chains short and jumped to answers. ## Why the Damage Persists Brain rot did not fully heal. The researchers scaled instruction tuning from 5,000 to 50,000 Alpaca examples and, separately, added clean continual pre-training on C4 web text. Both helped, instruction tuning more so, but gaps remained of 17.3% on ARC-Challenge with chain-of-thought, 9% on RULER and 17.4% on AdvBench. The authors read this as persistent representational drift: junk changed the model's internal representations, not just its answer format. A training-free fix showed limits too. Self-reflection did not help, while external critique from GPT-4o-mini reduced thought-skipping, a workaround rather than a cure. If you plan further training, our overview of [how to train your own LLM](https://www.cognativ.com/blog/how-to-train-your-own-llm/) explains why evaluating before and after each step matters. ![A junk-trained model goes through instruction tuning or clean pre-training; both paths end in partial recovery below baseline.](https://storage.ghost.io/c/cc/dc/ccdca949-0fe4-4119-8990-70a44ad75060/content/images/2026/10/llm-brain-rot-junk-data-3.png) More instruction tuning and clean continual pre-training both helped, but neither restored the model's original performance. ## Brain Rot, Model Collapse and Data Quality Research Brain rot adds to evidence that training data content shapes capability as much as volume. Shumailov and colleagues, in [Nature in 2024](https://doi.org/10.1038/s41586-024-07566-y?ref=cognativ.com), showed that models trained indiscriminately on model-generated content lose the rare tails of the original distribution, a process they named model collapse. Brain rot differs: its junk is human-written, and the harm appears as skipped reasoning rather than lost diversity. The positive side is well documented. The [FineWeb paper](https://arxiv.org/abs/2406.17557?ref=cognativ.com) reported that FineWeb-Edu, a 1.3-trillion-token educational subset, performed much better on MMLU and ARC. [DataComp-LM](https://arxiv.org/abs/2406.11794?ref=cognativ.com) found model-based quality filtering key to a strong training set. [Lee et al. (ACL 2022)](https://arxiv.org/abs/2107.06499?ref=cognativ.com) showed deduplication cut memorized output tenfold, and ["Textbooks Are All You Need"](https://arxiv.org/abs/2306.11644?ref=cognativ.com) reported a 1.3B coding model reaching 50.6% pass@1 on HumanEval with about 7 billion tokens of textbook-quality and synthetic data. Data quality works as a lever in both directions. ## Mitigating the Risk Mitigating brain rot means treating data quality as a first-class engineering concern and monitoring after every training update. ### Systematic Cognitive Health Checks The authors recommend routine cognitive health checks, akin to safety audits in other industries. In practice that means a fixed evaluation suite run before and after each update: chain-of-thought reasoning, long-context retrieval, safety red-teaming and, where relevant, behavioral probes, plus tracking how often the model skips reasoning steps. Running these checks within an [AI governance program](https://www.cognativ.com/ai/governance/) helps catch drift before it reaches users. ### Prioritizing Data Quality Over Quantity Data quality should be the cornerstone of AI training. Instead of hoarding massive datasets, curate them: filter by semantic quality, deduplicate, distrust engagement as a proxy for value, and record provenance. Our guide to [training an LLM on your own data](https://www.cognativ.com/blog/train-llm-on-your-own-data/) covers dataset preparation, and a structured [data analytics practice](https://www.cognativ.com/consulting/data-analytics/) makes quality scoring repeatable. ### The Role of Instruction Tuning and Ongoing Research Instruction tuning with higher-quality data partially reverses brain rot but cannot restore baseline performance, so prevention beats repair: keep a clean checkpoint, gate new data before training, and roll back if a health check fails. Stronger mitigation methods remain an open research question. ![A cycle: curate sources, filter and deduplicate, train the update, run a health check, then return to curation.](https://storage.ghost.io/c/cc/dc/ccdca949-0fe4-4119-8990-70a44ad75060/content/images/2026/10/llm-brain-rot-junk-data-4.png) Prevention beats repair: curate and filter data before each update, then run a cognitive health check before release. ## What the Study Does Not Show The authors call this a pilot study. Junk data came from one platform, the experiments used four small and mid-sized open models rather than frontier systems, M2 relied on an LLM judge, and training volumes were small. The study shows junk continual pre-training can damage capable models in a controlled setting; it does not show that today's commercial assistants are degrading. ## Frequently Asked Questions Can AI models really get brain rot? In a controlled study, yes. Continual pre-training on short, viral, low-quality posts lowered reasoning, long-context and safety scores in four open models. "Brain rot" is an analogy. What counts as junk data for an LLM? Short posts with very high engagement, or posts an LLM judge labeled as clickbait, conspiracy, misinformation, spam or superficial content. What is thought-skipping? The main failure mode found: junk-trained models cut short or skipped reasoning steps and jumped to answers. Can brain rot in AI be reversed? Only partly. More instruction tuning and clean pre-training helped but did not restore baseline scores. Is brain rot the same as model collapse? No. Model collapse comes from training on model-generated content and erodes rare data. Brain rot comes from human-written junk and mainly harms reasoning and safety. Does this affect ChatGPT, Claude or Gemini? Those systems were not tested. The result is a warning about training practices, not evidence that any commercial assistant has degraded. ## Conclusion: Data Quality Is a Safety Issue The brain rot study gives concrete shape to an old rule: garbage in, garbage out. Junk web text lowered reasoning, long-context understanding and safety in every model tested, raised dark traits, and left damage cleaner training could not fully undo. With research on model collapse, deduplication and filtered datasets, the evidence points one way: data quality decides how capable and trustworthy a model becomes. Curate training data, treat engagement-driven sources with suspicion, run cognitive health checks after every update, and keep a clean checkpoint. If you plan to fine-tune or continually train a model and want a data and evaluation plan that guards against this decline, [talk to Cognativ about your AI project](https://www.cognativ.com/contact/).