It depends on the use case. Example: I walk through the airport and an application scans my face and either calls security because I was recognized as a known terrorist or let's me pass because I'm awesome. The parent's point matters for what happens after. In the not a terrorist case the problematic system would retain that (it thinks) you walked through the airport. In a even more problematic system you now get ads because you were there.

Wait - wouldn't you _want_ the system to retain that information in your first case if you were mis-identified as a terrorist? I.e. you were recognized, but you were confirmed not a terrorist - so you won't be flagged later. Or maybe I misunderstood.

Otherwise, you run the risk of not updating the model to fix the false positive, and you might be flagged multiple times in the future.

I wouldn't want the system to keep information about me having been anywhere unless I'm a criminal. How would that information be helpful in updating the model. It already worked correctly.

Interacting with a system can take years. That's okay for privacy if the person is informed and has control, e.g. can request full deletion anytime.