LLMs can be “untrained” to forget specific data🧠, vital for legal and ethical compliance. However, this “forgetting” is often an illusion. A technique called quantization, which simplifies the model’s internal structure, can surprisingly restore this “unlearned” information 🤯. This poses challenges in balancing utility (the model’s overall performance) and security (preventing data recovery). While methods like gradient ascent ⬆️ and negative preference optimization ⛔ exist, they aren’t foolproof against quantization. Ideally, unlearning should be effective, preserve utility, and prevent recovery. Current methods struggle with this trifecta, highlighting the need for further research. The illusion of forgetting emphasizes the importance of understanding LLM limitations for building robust and reliable AI.