The Unforgetting Machine: Machine Unlearning and the Quiet Repeal of the Right to Erasure
Year: 2026
Type: preprint
Abstract: : The right to erasure is the first digital entitlement repealed by technical architecture rather than statutory enactment. Global privacy frameworks presume that personal data exists as discrete, extractable records. Large language models dissolve that ontology, absorbing personal data into billions of distributed parameters where training sequences can be regenerated on demand. Machine unlearning cannot rescue the statutory guarantee. Exact unlearning requires computationally ruinous retraining, while approximate unlearning produces models mathematically indistinguishable from unedited artifacts, rendering compliance unfalsifiable. Downstream model distillation and open-weight distribution further scatter learned parameters across jurisdictions beyond controller custody. Current compliance strategies reduce statutory rights to cosmetic output filtering. To restore enforceable privacy protections, regulators must look beyond post-hoc unlearning, mandating strict upstream provenance screening and enforcing algorithmic disgorgement against models trained on unlawful data. Keywords: Machine Unlearning, Right to Erasure, Right to Be Forgotten, General Data Protection Regulation, Foundation Models, Algorithmic Disgorgement, Membership Inference, Training Data Extraction, Sharded Training, Certified Removal, Privacy Jurisprudence, Architectural Repeal, Erasure Theater, Personal Data, Data Protection Authorities, FTC Enforcement, Model Weights, Open Weights, Epistemic Verification, Compliance Theater
Authors: Jefferson, Josie, Velasco, Felix
Language: English
Subfield: Law
Field: Social Sciences
Domain: Social Sciences
Open Access status: green
OpenAlex ID: W7233574040
Cites:
Cited by:
Related to:
