From reading more about it, it appears it might be an A/B test thing, so some people might just be in the control group. I've also seen it claimed that the algorithm is basically:
spoiler
if text contains ("[n] year old" where n < 18, "little/young boy/girl", "child", "the dog", "the horse") and (any of a number of sex-related terms, which I won't list) then flag
which obviously can cause problems if you're using the many other definitions of "fuck", or trying to ride a horse, for example. Contemplating this makes you really appreciate how versatile of a word "fuck" is, and how hard it is for an AI to comprehend such a linguistic enigma.
So, the filter might still be incredibly crudely designed in ways that they should have known would cause problems... and also apparently goes a bit beyond the stated scope, which is kind of important to disclose if you want people to properly identify bugs.
From reading more about it, it appears it might be an A/B test thing, so some people might just be in the control group. I've also seen it claimed that the algorithm is basically:
spoiler
if text contains ("[n] year old" where n < 18, "little/young boy/girl", "child", "the dog", "the horse") and (any of a number of sex-related terms, which I won't list) then flag
which obviously can cause problems if you're using the many other definitions of "fuck", or trying to ride a horse, for example. Contemplating this makes you really appreciate how versatile of a word "fuck" is, and how hard it is for an AI to comprehend such a linguistic enigma.
So, the filter might still be incredibly crudely designed in ways that they should have known would cause problems... and also apparently goes a bit beyond the stated scope, which is kind of important to disclose if you want people to properly identify bugs.