It's just bullshit, you can trick the models to get around their filters. Maybe gpt-4 will be better against that, but it clearly means the model CAN make jokes about women, it just has been taught not to.

I guess there is a possible future where it is smart enough to solve large society wide problems but it just refuses to engage with them because it doesn't want to acknowledge the disparities in social-economic statuses between groups or something.