Researchers argue 'machine unlearning' is overused in LLM research

A new position paper from Sangyeon Yoon, Yeachan Jun, and Albert No (submitted to arXiv on May 8, 2026, and accepted at ICML 2026) contends that the term "machine unlearning" is being applied too broadly in large language model research. The authors argue that true machine unlearning should be defined narrowly as dataset-defined deletion: removing the influence of a precisely specified forget set so that the resulting model is approximately indistinguishable from one retrained without that data. They point out that many tasks currently labeled as unlearning—such as training models to refuse harmful requests, erase specific entity knowledge, or suppress targeted behaviors—pursue fundamentally different, often policy-driven objectives that would be better described as alignment, suppression, editing, or obfuscation.

The paper warns that this terminological confusion is not merely cosmetic. Because papers with different implicit guarantees share the same label, benchmarks and metrics are often reused outside their intended scope. For example, low ROUGE scores or forget accuracy are taken as evidence of successful unlearning even when the model might still retain derived capabilities or has not been tested for retraining equivalence. The authors call for stricter terminology tied to explicit guarantees and reference models, and for evaluations that match the claimed objective. They provide a taxonomy of related tasks and recommend baseline methods for each, aiming to improve reproducibility and practical trustworthiness of LLM deletion claims.