The quiet race to make models forget

Every model remembers more than it should. A name pulled from a scraped forum, a face from a dataset nobody audited, a sentence a person asked to be forgotten years ago — all of it folded into billions of weights, impossible to point to and, for a long time, impossible to remove.

That last assumption is the one a small group of researchers spent the past eighteen months trying to break. They call the field machine unlearning, and the premise is deceptively simple: take a trained model and surgically excise the influence of specific data, as if it had never been seen.

Learning is addition. Forgetting is surgery. — An unlearning researcher

Why deletion is so hard

A neural network does not store facts the way a database does. There is no row to delete. Information is smeared across the whole system, entangled with everything else it knows. Pull one thread and the rest can unravel.

• Exact unlearning: retrain from scratch without the data — correct, but ruinously expensive.
• Approximate unlearning: nudge the weights to undo a sample's influence — cheap, but hard to verify.
• Differential privacy: never memorise the sample in the first place — prevention over cure.

What follows is a map of that surgery — the techniques that work, the ones that quietly do not, and the regulators now betting that the difference can be measured.