Machine Unlearning: The Lobotomization of LLMs

This commentary explains that while GDPR-style deletion requests are straightforward for traditional databases, they are extremely challenging for large language models because training data becomes deeply embedded in model behavior. It outlines machine unlearning as an active research area (techniques to de-emphasize or wrap models), notes the practical and benchmarking limitations of current approaches, and calls for efficient unlearning pipelines to meet legal, ethical, and operational needs.