Federated unlearning in AI: Structural tension between privacy rights and cybersecurity in global data governance regimes

Indigenous data governance frameworks like OCAP (Ownership, Control, Access, Possession) and Māori *kaitiakitanga* treat data as a collective resource requiring relational stewardship rather than individual erasure. Federated unlearning, by decentralizing data processing, superficially aligns with these principles but risks replicating colonial extraction by outsourcing control to technical elites. The absence of Indigenous knowledge in this debate reflects a broader pattern where Western legal and technical solutions are imposed without consultation, ignoring millennia of relational data practices. True data sovereignty would require co-design with Indigenous communities, not algorithmic band-aids.

The tension between privacy and surveillance is not new; it echoes 20th-century debates over wiretapping and the rise of the welfare state, where legal frameworks lagged behind technological capability. Federated unlearning mirrors historical attempts to 'decentralize' accountability, such as the 1970s Fair Information Practice Principles, which failed to curb corporate data monopolies. The 'right to be forgotten' itself has roots in European legal traditions dating back to medieval defamation laws, but its application to algorithmic systems reveals how outdated concepts are repurposed for digital contexts. This pattern suggests a cyclical failure to address structural power imbalances in data governance.

In India, the *Aadhaar* biometric system demonstrates how 'privacy-enhancing' technologies can be weaponized for exclusion, with marginalized castes and Muslims disproportionately targeted under national security pretexts. China’s social credit system, while framed as a governance tool, shows how decentralized data processing can enable authoritarian control under the guise of efficiency. Meanwhile, in Latin America, community-led data cooperatives in Brazil and Mexico are experimenting with federated models that prioritize collective consent over individual rights. These examples reveal that 'privacy' is not a universal value but a site of cultural and political struggle.

Federated unlearning is a machine learning paradigm where models are updated without centralizing raw data, theoretically reducing exposure to breaches. However, research shows that gradient leakage attacks can reconstruct sensitive data from model updates, undermining privacy claims. The scientific literature also highlights that differential privacy techniques, often paired with federated learning, introduce trade-offs between utility and privacy that are rarely discussed in policy debates. Additionally, the lack of standardized metrics for 'unlearning' efficacy complicates regulatory compliance, as enforcement becomes subjective.

Artistic responses to data extraction, such as Hito Steyerl’s concept of the 'poor image,' critique how digital artifacts circulate as commodities while erasing their origins. Indigenous storytelling traditions, like the Māori *pūrākau*, frame data as living knowledge requiring relational reciprocity, contrasting with the extractive logic of AI systems. The spiritual dimension of data—seen in concepts like the Buddhist idea of *anatta* (non-self)—challenges the Western obsession with individual ownership, suggesting that 'unlearning' might instead require collective forgetting of harmful narratives. These perspectives invite a reimagining of data not as a resource to control but as a relationship to steward.

Scenario modeling suggests that without robust governance, federated unlearning could lead to a bifurcated internet where privacy is a luxury for the Global North while the Global South faces intensified surveillance. The rise of 'unlearning-as-a-service' models may concentrate technical expertise in a few corporations, creating new forms of data feudalism. Conversely, if paired with open-source auditing tools and participatory design, federated unlearning could enable decentralized data cooperatives to challenge corporate monopolies. The long-term risk is that this approach becomes a fig leaf for continued data extraction, masking deeper structural issues in AI development.

Marginalized communities, particularly Black and Indigenous groups, are disproportionately subjected to data harms like predictive policing and algorithmic bias, yet their perspectives are excluded from technical debates about 'unlearning.' Survivors of domestic violence and LGBTQ+ individuals face unique risks when data deletion is framed as a universal solution, as their safety often depends on historical records. The technical elite driving AI governance are overwhelmingly from privileged backgrounds, ensuring that solutions reflect their worldviews rather than the needs of those most affected. True inclusion would require ceding decision-making power to affected communities, not just tokenistic consultations.