ML is very difficult for porn. You end up with so many false positives it almost becomes useless compared to human moderation. "Porn" is also a very context sensitive term. What is porn vs. nude art? Is it still porn if the people in it aren't "pink"? What about cultural differences? What is considered inappropriate in the US may not be inappropriate in Europe, etc. How much skin has to be shown for it to be porn? There are so many questions. I'm sure detecting porn accurately will be one of the hardest problems we'll overcome in computer vision, as outrageous as that sounds - because of the level of context that is required.

See also: SFW porn - where other images are super imposed on top of real porn. It's hilarious, but is it really SFW? You decide. http://www.reddit.com/r/sfwporn