Deleting a data point from an AI image generator does not mean it stays deleted.
A new paper looks at diffusion data-point unlearning, the process of removing a specific image or concept from a trained diffusion model. Most evaluations check whether the deletion worked right after it happens and stop there. But real systems process deletion requests one after another, repeatedly updating the same model. The researchers tested what happens over that whole sequence and found a failure mode they call sequential reappearance: a data point that is correctly forgotten right after deletion can resurface later, even though nobody reused the deleted data or deliberately retrained on it. They also found a technical tell - targets that later reappear show sharper local denoising-loss geometry right after deletion than targets that stay forgotten for good.
This matters because unlearning is the mechanism companies point to when they promise a data point is gone - for copyright takedowns, privacy requests, or removing harmful content. A one-time check after deletion has been the de facto standard. This paper says that check can pass while the underlying forgetting is unstable, and only shows up as broken after later, unrelated updates touch the same model. For anyone relying on unlearning as a compliance answer rather than a technical one, that is a gap worth knowing about before regulators or plaintiffs find it first.
The quiet admission here is that unlearning research has been grading itself on a test that does not match how production models actually get updated. A deletion that cannot survive the next deletion is not really a deletion - it is a delay.