There's deep tension in the current ask-for-forgiveness-free-for-all approach to acquiring data for model training. Will "open" models cause this tension to reach a breaking point?