Entangled Representations Worsen AI Model Unlearning Damage.

A long-standing hypothesis in AI interpretability is that when different pieces of knowledge are intertwined within a neural network's internal representations, it becomes harder to selectively remove specific information without harming other, unrelated knowledge. This phenomenon is known as "collateral damage" during model unlearning. This study provides direct experimental evidence to support this intuition. Researchers used a technique called Selective Gradient Masking to train a suite of language models with varying degrees of disentanglement between different knowledge domains (e.g., biology vs. non-biology). When applying standard unlearning methods, the models with more disentangled representations consistently showed a better balance between forgetting target information and retaining other knowledge. Specifically, the most disentangled models experienced significantly less collateral damage, demonstrating that representational entanglement is indeed a key factor in the difficulty of effective model unlearning.