Machine Unlearning

Erasure assumes a row you can find. A trained model does not have one.

The one unlearning method that comes with a proof costs a full training run, and the benchmark that defines "efficient" rules it out. Compliance teams are promising deletion anyway.