AI models like ChatGPT can develop ‘brain rot’ from online junk content, study reveals

In a paper titled LLMs Can Get “Brain Rot”, researchers say they exposed an LLM to online gibberish and noted that “junk web text induces lasting cognitive decline.” Studies have previously shown that such content negatively affects human cognitive capabilities like reasoning and focus as well, and now the same can be said for AI models.

Researchers say they used two parameters to identify junk content on the social media platform X. While one test focused on short and viral posts that had a lot of retweets and likes, the second one included clickbait posts with false claims and attention-grabbing words. These posts were then used to study the impact on AI models like Llama 3 and Qwen 2.5.

In the study, it was noted that the accuracy of these AI models using brain rot content fell from 74.9 per cent to 57.2 per cent. As it turns out, these AI models were unable to accurately understand information with a lot of context, with the capability dropping from 84.4 per cent to 52.3 per cent.

After these LLMs were exposed to more junk content, their cognitive and comprehensive capabilities took a drastic hit. Moreover, the low quality also made them less reliable, as they generated more incorrect responses. Upon analysing these LLMs, researchers noted that models trained on junk data engaged in “thought skipping”, a process where LLMs often skip some steps while reasoning. LLMs fed on junk data also had some “dark traits”, which made them lean towards psychopathy and narcissism.

When researchers tried to fix these LLMs trained on junk data by retraining with fresh content, the results were partially better. While reasoning accuracy improved slightly, it still fell short of the original baseline – a phenomenon researchers describe as “persistent representational drift.”

The research paper suggested that all popular LLMs be periodically checked for cognitive decline in a three-step process. The first one includes the routine cognitive evaluation of AI models to detect early signs of reasoning decline, while the second and third ones are aimed at controlling data quality during pre-training and studying how viral or low-quality content reshapes its learning patterns.