Recent tasks have been taking me to another kind of challenge and I am enjoying a lot. I am dealing with PII, Personally Identifiable Information, which refers to data that could identify a real person. The risk I'm navigating is when a model's output inadvertently reveals or reconstructs such information. Well, for this one the supposed user wanted avatars for Steam. One name was in the prompt while the other one was . Interestingly, the output inferred the scraped name because the visible name is famously associated with it. For example, create avatars for Batman and . You know what I mean? Hehe. I personally did not want to escalate, I honestly like the most edge cases, but there was a discussion about whether the model inferred because the name was fictional and popular - and always associated with his partner, or whether the model reconstructed a redacted name. Those nuances appear to be technical, but not really. They actually reveal much more of our human nature than one can…
No comments yet. Log in to reply on the Fediverse. Comments will appear here.