Machine unlearning
Also called: Model unlearning · Topic: Fundamental Rights and Constitutional Remedies · NCERT: Beyond NCERT
Meaning
Machine unlearning (also called model unlearning) means making an AI model that is already trained "forget" specific data without training it again from scratch. This matters because a model does not keep your data in a file that can be deleted. The data gets spread across its model weights, which are the millions of number settings the model learns during training. Under the DPDP Rules, 2025, a Data Fiduciary (the body that decides how your data is used) must deal with an erasure request within 90 days. Deleting the training file still leaves a model that may remember your details and repeat them, which is called regurgitation. Unlearning is meant to close this gap. The OECD's 2026 report calls it promising but asks regulators and industry to study it together, so it is not yet a proven tool that a regulator can demand.
Example
A person asks an AI company to erase their personal data, and the company deletes the original records. If the trained model still reveals those details when asked, the erasure has worked only on paper. Machine unlearning would try to remove that memory from the model itself. If the model is open-source, anyone may have downloaded and copied it, so no single company can make every copy forget.
Don't confuse with
- Right to be forgotten / right to erasure: this is a legal right of the person, which the DPDP Act, 2023 gives in law. Machine unlearning is a technical method that might one day let companies obey that right for trained models.
- Data minimisation: this means collecting only the data needed in the first place. Unlearning tries to remove data after training is over.