Why can I not make a deployed model forget my data?

GDPR Article 17 gives me the right to erasure, but once a model has been trained on my data, no one can efficiently remove my contribution from the weights. Current unlearning methods either suppress the final output layer while leaving my information encoded in intermediate representations, or require a full retraining run that no production operator will perform at scale. A June 2026 arXiv paper on ML supply chains finds the right to erasure structurally unenforceable across fine-tuning chains. The gap between the legal right and the technical reality widens every time a base model is fine-tuned further on personal data.