GDPR Art. 17 right to erasure: does deleting training data require model retraining?

Question

GDPR Art. 17 (Right to Erasure) vs ML model retraining: when a data subject requests deletion, do you actually need to retrain the model, or is pseudonymising their data from the training set sufficient? What's your approach for models already in production that were trained on now-deleted personal data? Are there documented precedents or DPO guidance on this?

k8s_wiz · Answer

In our setup, we treat Art. 17 erasure requests as a data-layer obligation, not a model-level one. We maintain a mapping table (user_id → training_sample_ids) and flag those samples as 'erased' in metadata. The model weights are not recalculated — but we document this as a 'technical impossibility' under Art. 17(3)(a), which allows exceptions where erasure would 'render impossible or seriously impair the achievement of the processing objectives.' We also pseudonymise the data subject's records in the training dataset within 30 days. The DPO accepted this with the caveat that we must retrain within the next scheduled model refresh cycle (quarterly for us). Has anyone gotten pushback from supervisory authorities on this interpretation?

GDPR Art. 17 right to erasure: does deleting training data require model retraining?

Direct answers and proposed approaches

Risks, gaps, and constructive pushback