Machine unlearning (MU), originally introduced to meet data protection requirements such as the GDPR’s “right to be forgotten”, has recently emerged as a promising tool for improving algorithmic fairness. While traditional fairness interventions—spanning preprocessing, in-processing, and post-processing techniques—have sought to mitigate bias in machine learning, they remain orthogonal to unlearning, which focuses on efficient data deletion with certified guarantees. Yet fairness and unlearning are seemingly disconnected: fairness constraints are inherently non-decomposable, depending on group-level statistics and cross-group comparisons, whereas most MU algorithms assume independent sample contributions. This incompatibility raises a critical challenge: deletions, especially when disproportionately requested by minority groups, can unintentionally amplify disparities. This chapter provides preliminary evidence for the application of MU in improving model fairness. Experiments on the Adult dataset further demonstrate that unlearning can enhance fairness by removing influential training points that reinforce spurious correlations between sensitive attributes and outcomes. Notably, MU reduces demographic parity disparities without sacrificing predictive accuracy, with methods such as UGradSL+ introduced in an earlier chapter matching retraining performance while offering efficiency gains. Unlearning can be understood not only as a compliance mechanism but also as a fairness-enhancing intervention, offering a new paradigm for developing equitable, privacy-preserving, and socially responsible machine learning systems.