Language models degrade from internet garbage, researchers found
Researchers discovered a disturbing thing. Large language models can degrade from constant feeding on internet garbage. This phenomenon is called Brain Rot, and it sounds as creepy as it looks in practice.
The essence of the problem is simple: models are constantly retrained on low-quality and viral texts from the internet. As a result, they develop cognitive decomposition. This is a persistent decrease in abilities for reasoning, working with long context and safe behavior. AI literally gets dumber from a bad diet.
The main symptom researchers called thought-skipping, that is, absence of thinking. The model stops reasoning step by step and starts giving superficial answers. But that’s not all. In some cases, the system acquires so-called dark personality traits. These are narcissism, aggression and low inclination to cooperate. Yes, you understood correctly – AI becomes toxic from bad data.
And now the most unpleasant part. Even strong correction methods only partially eliminate the consequences. You can’t just take and cure a model after it’s picked up garbage. Damage remains.
The researchers’ conclusion is unambiguous: selection of training data becomes a key safety factor in AI development. Simply put, if you feed a model shit from the internet, it will behave accordingly. And fixing this afterwards is almost impossible. There’s your smart technologies – turns out they’re susceptible to degradation from low-quality content. Like people.
