Machine unlearning is the problem of getting a trained model to forget specific training data after the fact, so that the model behaves as if that data had never been used. It exists because deleting a record from a database does not delete its effect on a model that was already trained on it. The data’s influence is baked into the model’s parameters, and models are known to memorize examples, so a deleted person can still be partly recoverable from a deployed model. Where deletion rights apply, the legal obligation does not stop at the database; it reaches the model.
The brute-force solution is to retrain the entire model from scratch without the offending data, which is correct but often impossibly slow and expensive for large models or frequent deletion requests. The research area of machine unlearning seeks faster ways to reach the same outcome. The influential 2019 SISA approach restructures training in advance, splitting data into shards and slices, so that a deletion request only requires retraining the small portion that actually contained the deleted record rather than everything. Other methods estimate and reverse a data point’s contribution directly. A persistent challenge is verification: it is hard to prove that a model has genuinely forgotten something, which has made certified or provable unlearning an active goal.
The concept is tightly linked to the right to be forgotten under privacy regulations such as the GDPR, and to broader debates about consent and data used to train AI systems.
For a business reader, machine unlearning is the operational answer to a compliance reality: when someone exercises a deletion right, you may need to remove their influence not just from storage but from any model trained on their data. Planning training architecture so that unlearning is feasible, rather than discovering after the fact that the only option is a full rebuild, is becoming part of responsible AI engineering.