They also trained a model on what qualifies as a "sensitive" prompt, which means:

For a given prompt, the odds that it is considered sensitive is probabilistic

Since it's an input outside if the prompting part, they can probably tune that specific aspect independently in response to controversial usages