I spend way too much time on social media, and I already worry about what it does to my attention span. Now, researchers at Texas A&M University say the same kind of content can also mess with AI models.
The study, from assistant professor Dr. Zhengzhong Tu and his team, has been accepted to the 2026 Conference on Language Modeling. The team took 1 million public tweets from 2010 and sorted out the “junk.” A post counted as junk if it was short (under 30 tokens) but popular, with more than 500 combined likes, retweets, and replies. They also had GPT-4o mini flag posts that relied on hashtags, all-caps words, urgent language, or trivial topics.
Next, they trained four models, including Llama3 8B and three Qwen models, on different amounts of junk and tested them on reasoning, memory, ethics, and personality.
So what happens to an AI on a junk diet?
The models started skipping steps. Sometimes they gave an answer without reasoning through the question at all, which the researchers call thought-skipping. A post’s popularity had a bigger effect on reasoning, while its length did more damage to memory and understanding.
Some of the models also lost their moral compass. In one test, a model that had earlier refused to explain how to ruin someone’s reputation suggested spreading lies and gossip after training on 100% junk data. It did add a note that this was unethical, which is not much comfort.
What surprised me most is how stubborn the damage was. Retraining with cleaner data didn’t fully fix it.
Why should regular users care?
Tu points out that AI already generates far more content than humans do. If models learn from junk, they can produce more junk, which then feeds both our timelines and future models. In other words, AI and humans could end up dragging each other down.
As Shuo Xing, one of the lead authors, put it, “AI health and human cognitive health are not separate issues.” There’s also no easy way to check whether ChatGPT or Gemini were trained on junk, since their training data isn’t public. Hence, Tu wants AI development to be more open so researchers can actually inspect the data.
RELATED COVERAGE: