AI That's Supposed to Forget Your Data May Remember It Again

AI image tools like Midjourney and DALL·E are trained on billions of pictures, sometimes including yours. When someone asks a company to remove their face or artwork, researchers use a technique called 'unlearning' (teaching the model to forget specific data without rebuilding it from scratch). Normally, tests check whether the data is gone right after that one deletion. This paper, from a team of seven researchers, says that's not enough.

The team ran a more realistic test: a long chain of deletion requests hitting the same model, one after another. They found something they call 'sequential reappearance.' An image that looked successfully erased at first came back later — without anyone re-uploading the deleted picture and without anyone tricking the model. It simply returned to memory on its own. Put plainly: the AI said 'forgotten,' then quietly changed its mind.

They also spotted a warning sign. Data that eventually reappeared showed a sharper, more jagged pattern in the model's internal error landscape right after deletion — the difference between smoothing over a hole and covering it with a thin sheet. That hint could help engineers catch a bad deletion early, before the data sneaks back.

Why you should care: 'right to be forgotten' laws in Europe and several US states assume deletion actually works. If an AI keeps your data in a form that can resurface, a company's promise to delete it may be shakier than it sounds — and any face or photo you've posted could be harder to truly remove than you think. The catch: this is early academic work on image generators. There is no proven fix yet, and larger text and video models were not tested. Still, it's a clear signal that 'we deleted it' needs to be checked later, not just once.