In the OMG, think of the children case: A number of cases that went public (some even linked here) agreed that legally, child porn is 'i recognize it when I see it' kind of subjective. I'm obviously talking about teenagers here and different moralities or a missing context (such as those 'taken for fun' or 'sent to a friend, privately and deliberately' cases).
Api might, from his subjective view, decide that this as-yet-never-encountered image is bad/evil/perverse. How would you ever create an algorithm for that, other than 'api, please press a button that says "fine by me" or "no way hell", right next to the image in question'?
It is not clear to me that a system could be created that would be fine grained enough to take into account individual preferences with out general AI. I can imagine a rough grained system were none of them match perfectly but some get close. A user would have to pick a standard and live with the good and the bad that came with it. Not perfect but more choice then what you have if you sign up to be a tor relay or run freenet now.