AI-pref is still in limbo/derailment state at IETF best i can tell, so i would recommend to wait to see where that data model lands and don’t make something out of alignment with it. and yeah at the lexicon level probably makes sense.
i have been ruminating on how an iscc field in a lexicon would work, for content whose author(s) want its authorship, licensing, and/or ai-prefs to be “findable” (vector/perc hashing) across edits and metadata stripping… seems to be the same metadata pattern, and there too metadata at the lexicon/MASL level makes the most sense imho