Glossary · Term

machine unlearning

← all terms

Definition

Plain language

Making a trained AI forget something specific — a book, a person's data — without retraining it from scratch.

As stated in the literature

The task of removing the influence of designated training data from a model; post-hoc methods suppress content via fine-tuning but are brittle (recoverable in a few relearning steps), motivating architectures where unlearnability is built in and removal is structural rather than cosmetic.

Also called: unlearning, unlearnable

Why it matters: It promises a practical way to honor data-removal requests, but shallow methods can leave the 'forgotten' content easy to recover, so the removal needs to be genuine.

For example, if a model was trained on a book the author wants removed, machine unlearning aims to strip out that book's influence without rebuilding the model from scratch.

Heard on the show

“… setup — a line saying "we want to determine whether you can be deployed," which clashes with the unlearning threat — and the sandbagging drops to about twenty-three percent. …”
Episode 174 — When the AI 'Schemes,' It's Usually Just Lazy or Confused

Mentioned in 2 episodes

  1. 174
    When the AI 'Schemes,' It's Usually Just Lazy or Confused
  2. 145
    Building Forgetting Into a Language Model With One Extra Line of Code

Related terms