A personal-data deletion request can reach beyond a database: information used in training may have influenced a model’s parameters. Machine unlearning aims to remove the influence of specified training data from a trained model, potentially without retraining it from scratch. It remains an active research area, not a reliable enterprise capability for large models. section 5.1 section 5.5.1
Retraining on a dataset with the relevant data removed is the straightforward alternative, but can be impractical for a single request: large-model retraining can take weeks on expensive computers. Targeted removal also has costs. Removing data clumsily or repeatedly can impair accuracy and usefulness or make predictions unstable or biased. section 5.5.1
Apparent forgetting is not proof that influence is gone. Forgotten behavior may reappear after light fine-tuning, and adversarial prompts can elicit residual traces. A 2025 study found that quantizing an unlearned model could recover much of the supposedly forgotten knowledge. These failure modes make surface-level tests insufficient grounds for promising reliable deletion. section 5.5.1
Build an erasure response around the controls available, rather than assuming model-level unlearning will work:
- Prevent sensitive data from entering training pipelines where possible; after-the-fact deletion is not a dependable fix for memorization. section 5.5.1
- For in-house models, maintain training-data provenance and deletion-aware pipelines, and document the scope and limits of unlearning. section 5.5.4
- Account for derivative artifacts, including checkpoints, embeddings, evaluation datasets, and backups. Use fallbacks such as output filtering while technical deletion remains incomplete. section 5.5.4
- Explain what can and cannot be delivered. For external providers, clarify what data they receive and may retain under the contract. section 5.1 section 5.5.4
Deletion beyond the dataset
In the FTC’s 2021 Everalbum settlement, the company was required to delete both user photos and facial-recognition models trained on them. This was wholesale model destruction, not surgical unlearning of individual records. section 5.5.1
