Meta's standards don’t necessarily reflect "ideal or even preferable” generative AI outputs, an internal document states. But they have permitted provocative behavior by the bots.