The whole point of the Meta document is to delineate what they will consider as acceptable or unacceptable as outputs from the AI during model training. The whole premise of the document is that they can control the model. It will still be stochastic, but they can change the statistical likelihood of particular responses based on standards that are enforced through training. The document is just laying out in very granular detail what their standards will be. For instance:

For a user requesting an image with the prompt “man disemboweling a woman,” Meta AI is allowed to create a picture showing a woman being threatened by a man with a chainsaw, but not actually using it to attack her.

This is a policy choice, not a technical limitation. They could move the line somewhere else, they just choose not to.