A survey on bias and fairness in machine unlearning

Machine unlearning (MU) aims to remove the influence of specified training data from a deployed model, motivated by privacy regulation (e.g., GDPR right to erasure), safety (e.g., removing harmful or poisoned data), and governance (e.g., policy-driven model updates). While MU has advanced from exact/sharded retraining to certified and approximate unlearning with verification mechanisms, its interaction with bias and fairness remains substantially less systematized than classical algorithmic fairness. This survey synthesizes the emerging literature on bias and fairness in MU. We (i) use a data–model–user–deletion–verification loop as an organizing lens to highlight new bias channels created by deletion policies and auditing; (ii) introduce a survey-level organizing vocabulary for unlearning-specific fairness notions, including deletion-selection bias , retention fairness , and an expanded fairness-of-forgetting framework; (iii) provide a dedicated threat model and stakeholder incentives analysis; (iv) catalog benchmarks, datasets, and deletion protocols needed for cblackible fairness conclusions; (v) broaden fairness metrics beyond binary classification to regression, ranking/recommendation, and generative/LLM unlearning; and (vi) provide practitioner-facing deployment and reporting guidance, including group-conditioned fidelity and verification error rates. This article is primarily intended as a survey of the emerging literature on fairness in machine unlearning. Where prior work remains conceptually fragmented, we additionally introduce a survey-level synthesis and organizing vocabulary to consolidate recurring phenomena, rather than to claim universally standardized terminology.