Position: The Term "Machine Unlearning" Is Overused in LLMs
06:00 · June 29, 2026 · arXiv cs.AI RSS
Large language models increasingly face demands to "forget" training data, knowledge, or behaviors due to regulatory deletion obligations, copyright/licensing disputes, and safety or product-policy requirements. This position paper argues that machine unlearning is overused as a term in LLM research and should be reserved for dataset-defined deletion: removing the training influence of a precisely specified forget set such that the resulting model is approximately indistinguishable from retraining without that data. We contend that many tasks currently labeled "unlearning" (e.g., refusal for harmful requests, entity/knowledge removal, or targeted suppression) pursue different, often policy-dependent objectives and therefore require different terminology and baselines (e.g., alignment, suppression, editing, obfuscation). We further argue that this confusion is not cosmetic: because papers make different implicit guarantees under the same label, metrics and benchmarks are frequently reused outside their intended scope, rewarding surface-level non-disclosure (e.g., low ROUGE/forget accuracy) even when retraining-equivalence is not tested and derived capabilities remain. We conclude by calling for stricter terminology tied to explicit guarantees and reference models, and for evaluations that match the claimed objective.
Summary
Large language models now face frequent demands to remove the influence of specific training data, driven by privacy regulations such as the right to be forgotten, copyright disputes, and safety or licensing constraints. A new position paper contends that the term “machine unlearning” has become diluted in LLM research because it is applied to objectives that differ sharply from its original technical meaning.
In the classical formulation, machine unlearning is a dataset-defined deletion task. Given a training set D and a precisely specified forget set F, the goal is to produce a model whose behavior is approximately indistinguishable from one obtained by retraining from scratch on D excluding F. Success is therefore measured against an explicit retraining baseline rather than against any particular output policy.
Many recent papers instead label as unlearning tasks such as inducing refusal of harmful queries, suppressing named entities, or editing factual knowledge. These interventions typically pursue policy-dependent behavioral constraints rather than removal of training influence. When the same label covers both dataset-defined deletion and policy-driven suppression, evaluation metrics become mismatched: measures such as reduced ROUGE scores or lower question-answering accuracy can reward superficial non-disclosure while leaving derived capabilities intact and without verifying equivalence to a retrained reference model.
The authors therefore advocate reserving “machine unlearning” for methods that target retraining indistinguishability and adopting separate terms—alignment, suppression, editing, or obfuscation—for other objectives. They further recommend that benchmarks explicitly state the intended guarantee and include reference models trained without the forget set, so that reported results reflect the actual claim being made rather than surface-level output control.
Why it matters
This paper is highly relevant for Dutch AI researchers and compliance officers dealing with GDPR's 'right to be forgotten' and the EU AI Act. By clarifying the distinction between true machine unlearning and mere suppression, it provides a crucial framework for developing legally compliant and transparent LLMs.
