Machine unlearning and the politics of algorithmic forgetting: Three logics of epistemic reconfiguration

Machine unlearning has emerged as a technical response to demands that AI systems remove the influence of unwanted data. Drawing on critical readings of technical literature and theories of cognitive assemblages, archives, and forgetting, in this article we examine the technical promise of machine unlearning as an ethico-political practice of epistemic reconfiguration. We identify three interwoven logics: influence mitigation through parameter reweighting, embedding-space reconfiguration, and output suppression. Across these logics, information is attenuated, relocated, or rendered inaccessible rather than made nonexistent. These operations redistribute what models can retrieve and express, with consequences not only for privacy, accountability and bias, but also deeper questions about justice, accountability and the ethics of cultural memory. We argue that machine unlearning should therefore be understood not as neutral technical deletion but as a form of world-unmaking that reorganizes knowledge and ignorance. Cultural theories of memory, archival silence, and critical unlearning help expose the political stakes of who decides what becomes reachable, suppressed, or forgotten in AI systems.