"Machine Unlearning" Term Overused in LLMs, Needs Stricter Definition.

A recent position paper argues that the term "machine unlearning" is being excessively and inaccurately used within the field of large language model (LLM) research. The authors contend that true machine unlearning should be strictly defined as the process of removing the influence of a precisely specified dataset from a model, to the extent that the resulting model is virtually identical to one that was never trained on that data in the first place. This rigorous definition is crucial for ensuring clear guarantees, especially in contexts like regulatory deletion obligations or copyright disputes. The paper highlights that many tasks currently categorized under "unlearning"—such as making an LLM refuse harmful requests, removing specific entities or knowledge, or targeted suppression of information—do not meet this strict definition. Instead, these objectives are often policy-dependent and should be more accurately termed as alignment, suppression, editing, or obfuscation. The authors warn that this terminological confusion is not merely semantic; it leads to the misuse of metrics and benchmarks, where surface-level non-disclosure might be rewarded even when the fundamental goal of retraining-equivalence is not achieved. They advocate for stricter terminology linked to explicit guarantees and reference models, alongside evaluations that precisely match the claimed objective.