It feels like any time I have a conversation about attributing data used to train AI models, the completely understandable impulse to want attribution starts to break when confronted with some practical implementation questions.