To Be Forgotten or To Be Fair: Unveiling Fairness Implications of Machine Unlearning Methods

The right to be forgotten creates a fairness dilemma: While people have a legitimate interest in having their data removed from AI systems (the "right to be forgotten"), the methods used to erase this data can inadvertently create new fairness problems, meaning that solving one ethical concern may introduce another.

Different unlearning methods have different fairness impacts: The study tested two machine unlearning approaches (SISA and AmnesiacML) and found they performed differently in terms of fairness. Specifically, SISA maintained better fairness than other methods when data was deleted unevenly across groups, suggesting that the choice of deletion method matters significantly for ethical outcomes.

How data is deleted matters more than whether it's deleted: The research found that uniform deletion of data (removing equal amounts across all groups) didn't necessarily harm fairness, but non-uniform deletion (removing more data from some groups than others) created fairness risks—highlighting that the pattern of deletion is as important as the act itself.

Practitioners need guidance on fairness trade-offs: The paper identifies this as an important gap in software engineering practice, suggesting that companies adopting unlearning methods need better tools and frameworks to understand and navigate the potential trade-offs between respecting individuals' right to be forgotten and maintaining fair AI systems.