Position: The Term "Machine Unlearning" Is Overused in LLMs

Short summary

This position paper argues that 'machine unlearning' is overused in LLM research and should be strictly defined as dataset-level deletion achieving approximate retraining equivalence. Many tasks labeled 'unlearning'—including refusal, knowledge removal, and alignment—pursue different policy-dependent objectives requiring distinct terminology. The paper calls for stricter definitions and evaluation metrics matched to claimed objectives rather than repurposing benchmarks (ROUGE, forget accuracy) across incompatible contexts.

• •Machine unlearning term is overused; should be narrowly defined to dataset deletion with retraining equivalence
• •Refusal, alignment, and knowledge removal are distinct objectives requiring separate terminology and metrics
• •Current benchmarks reward surface-level non-disclosure without validating true unlearning guarantees

Generated with AI, which can make mistakes.