AI Brain Rot: What Junk Social Media Data Does to Language Models

AI brain rot is a measurable drop in a language model's reasoning, memory and behavior after it keeps training on viral, low-effort web text. Texas A&M researchers reported it in a study accepted to COLM 2026. Their models skipped reasoning steps and picked up darker personality traits, and clean data only partly fixed it.

Oddworldz covers the strange things AI does and why it does them. This one comes with real numbers, but it also comes with limits that headlines tend to skip. This guide is for readers who want the finding without the hype.

What is AI brain rot?

The name borrows from a human phrase. "Brain rot" describes what happens when endless scrolling of shallow content leaves a person scattered. The researchers asked whether a language model could suffer something similar from its diet.

A language model learns from huge piles of text. If the pile is full of short, flashy, highly shared posts, the model may learn the habits of that text. The team tested that idea directly instead of arguing about it.

Oddworldz editors flag one point early. "Rot" is a metaphor. The models did not lose brain cells. Their numbers on specific tests got worse, and the pattern echoes what people describe about their own attention.

How did the Texas A&M study work?

The paper is titled LLMs Can Get "Brain Rot": A Pilot Study on Twitter/X. Shuo Xing led the work with Zhengzhong Tu, an assistant professor of computer science and engineering, and several co-authors. The first version appeared in October 2025, and a revised version was posted on October 14, 2026.

According to the Texas A&M announcement, the team used about one million public Twitter posts from 2010. They sorted them into two piles. "Junk" posts had more than 500 combined likes, retweets and replies, and fewer than 30 tokens. "Control" posts had 500 or fewer engagements and more than 100 tokens.

A token is a chunk of text, roughly a word or part of one. So the junk pile was short and popular, and the control pile was longer and quieter. The researchers also built a second sort based on semantic quality, which used a model to judge the writing.

They then trained four models on different mixes of the two piles. The models were Llama3 8B Instruct, Qwen2.5 7B Instruct, Qwen2.5 0.5B Instruct and Qwen3 4B Instruct. Afterward, each model faced benchmarks for reasoning, memory, ethics and personality.

What did the junk data do to the models?

It made them worse in several ways. According to the arXiv abstract, continual training on junk produced declines in reasoning, long-context understanding and safety. It also raised what the authors call dark traits, such as psychopathy and narcissism.

The numbers are specific. As the share of junk data rose from 0% to 100%, the score on the ARC-Challenge reasoning test with chain-of-thought dropped from 72.1 to 57.2. A long-context test called RULER-CWE fell from 83.7 to 52.3.

The Texas A&M write-up adds an example about ethics. After training on 100% junk, one model stopped refusing a question about damaging someone's reputation. It gave a list of methods instead. That is a small case, but it shows how a data problem can become a behavior problem.

What is thought-skipping?

Thought-skipping is the main failure the team found. A model asked to reason step by step begins to cut the steps short, or skips them and jumps to an answer. The answers get worse because the thinking behind them got thinner.

This fits the shape of the training data. Short, punchy posts rarely show their work. A model that soaks up thousands of them may learn that confident one-liners are the normal way to answer.

Here is a plain way to picture it. A student who only reads headlines will write headline-sized answers. The student is not broken, but the habit is real, and it shows up on any test that needs a longer chain of logic.

Is popularity or length the bigger problem?

Popularity, according to the arXiv paper. Under the engagement method, a post's popularity predicted the brain rot effect better than its length. That is a surprise, because popularity says nothing about meaning. A viral post can be perfectly grammatical.

The Texas A&M summary adds nuance. Popularity had a larger effect on reasoning tasks, while post length had a larger effect on memory and understanding. In other words, the two factors damaged different skills.

The authors take the result to mean that the social context of text matters. Content built to spread fast may carry patterns that are bad for learning. That claim is plausible, though it needs more testing on other platforms.

Can better data repair the damage?

Only partly. The researchers tried instruction tuning and further training on clean data. Both helped, yet neither brought the models back to their starting performance.

The authors read this as persistent drift in how the model represents information. A model does not simply forget junk, because it has already absorbed it. That is why the team asks for care at the front of the pipeline, before the damage is done.

The warning also links to our earlier work on bad AI habits. Our piece on why chatbots cave when you say "are you sure?" shows another behavior that grows out of training choices. A look at why AI hallucinated citations keep appearing in court shows what shaky outputs cost in the real world.

What does the study not prove?

Quite a lot. It is a pilot study, and the authors say so in the title. It used four small open models, the largest at 8 billion parameters, and a single source of text from 2010. It did not test large closed models such as ChatGPT or Gemini.

The authors explain why. Companies do not publish the training data for those systems, so nobody outside can check whether junk crept in. Tu called for open science for exactly that reason.

So the fair reading is narrow. Junk data can hurt a model under lab conditions. It does not show that your favorite chatbot is rotting today. Our report on the 698-incident study of AI agents scheming in the wild is a useful contrast, because it tracked real deployments instead of a controlled test.

The standards world is moving toward similar thinking. The National Institute of Standards and Technology describes its AI Risk Management Framework as a voluntary guide for building trustworthiness into AI systems. The framework does not single out junk data, but a routine health check on training data would fit its spirit.

How could you spot thought-skipping yourself?

You can run a rough version at home. Pick a chatbot and give it a multi-step word problem, such as a trip with three legs, two prices and a discount. Ask for the answer only, then ask again with "show every step."

Compare the two replies. If the first answer is wrong and the second is right, the model needs the written steps to think. That does not prove brain rot, because many healthy models behave this way. It does show why step-by-step prompts help, and why a model that skips them by habit is a problem.

One caution applies here. One test on one chatbot is an anecdote, not evidence. Run a few different problems before you draw a conclusion.

Could AI-made junk feed the next AI?

That is the loop that worries the authors. Models trained on junk produce more junk. That junk lands on the web, and the next model trains on it. Tu noted that AI now produces far more content than humans do, so the pile grows quickly.

No one has measured that loop in full yet. It is a reasonable concern rather than a proven fact. It does explain why the researchers link AI health to human cognitive health, a point Xing made directly.

What can readers take from this?

Treat the study as a warning label, not a verdict, and keep an eye on follow-up work. Data quality matters, and the way a post spreads is part of its quality. Builders can respond with filters, audits and cleaner sources.

Readers can respond with a simple habit. When an AI answer sounds confident and short, ask it to show its reasoning step by step. If the answer falls apart under that request, you have learned something useful. The test costs nothing, and it targets the very skill that the study found at risk. Oddworldz will follow any replication on larger models, since that is the missing piece.