Right to Be Forgotten

This concept is the legal door that leads to the technical room of machine unlearning. Understanding it first means you know why companies must build systems to forget data, and what they must prove to regulators. It unlocks further study of: Exact Unlearning, Approximate Unlearning, SISA Sharded Training, Federated Unlearning, Utility Preservation, Unlearning Verification, and Membership Inference — all of which you will need for real-world AI compliance work.

Think of a warehouse full of cardboard boxes, each box containing a copy of a person's information. The law says that when someone asks for their data to be deleted, you must destroy every copy — not just the one in the front office, but the ones in the basement, the backups, the ones accidentally shipped to another branch. Now imagine the warehouse isn't boxes but a giant recipe book. The recipe for a model — its weights — was influenced by thousands of ingredients, including one person's data. You can't just rip out one ingredient and still have the same recipe; the cake has already been baked. The 'Right to Be Forgotten' extends to this baked cake: you must somehow make the cake as if that ingredient had never been added. That's the hard problem. The law doesn't care how hard it is — it just sets a deadline.

Imagine you are a chef who has prepared a large pot of soup for a banquet. The soup is the trained model. Each ingredient is a piece of personal data. A guest suddenly says, 'I didn't give you permission to use my secret spice blend — remove it.' You can't fish out the spice from the simmering pot; it has dissolved and changed the whole flavour. Your options: (1) throw out the whole pot and start over from scratch with every other ingredient — this is full retraining, the slow but correct baseline; (2) try to add another ingredient to mask the flavour — this is approximate unlearning, fast but not guaranteed to truly remove it; (3) from the beginning, cook each spice in its own tiny pot and only mix them at the end — this is SISA sharded training, where each ingredient touches only one small pot, so removing one spice means throwing away only that tiny pot and re-mixing. The analogy breaks down because in cooking you can taste the result, but for AI we can't 'taste' whether the model truly forgot — we need clever tests to check.

The legal right to have personal data erased, which extends to models trained on that data and not merely to databases.

Before this page, you have operationalised the idea of a 'model' as a collection of numbers (weights). This concept connects that idea to the legal world. It is the reason the technical problem of machine unlearning exists. It sits directly above Exact Unlearning (removing influence completely), Approximate Unlearning (removing influence approximately), and Deletion Requests (the process of handling a request). It also connects to Utility Preservation (the tension between forgetting and keeping the model useful) and to Verification (the hard problem of proving forgetting happened).