I don't envy @CobaltVelvet's attempt to moderate their instance. Like, I recognise how important it is, and getting user buy-in is actually one of the best way to do it.
But wrangling Internet randos — and I count myself as one of them — into buying in to your policy and then dealing with the administrative work of ensuring that the policies are followed... is grueling, thankless work.
Hence that decision of wanting an instance for friends only. I'm guessing calling out friends will be easier.
@tariqk @CobaltVelvet Random thought just occurred to me.
If there's a way to flag a toot as inappropriate, and that is fed to some list that admin has access to, then the hand-moderation of those toots can be fed in to a neural network which runs in parallel but doesn't affect moderation results until it's capable of matching the hand-selected moderation outcome to a certain degree of accuracy.
FWIW, I don't know much about NNs, but this seems to be the kind of fuzzy task they're great at...
@CobaltVelvet @orielle From what I know of neural nets, you're going to need to have a pretty big dataset, the dataset better not be polluted with implicit bias from the data collectors and people providing the rating of offensiveness, and, from my understanding of a couple of recent reports, might actually be fooled by some trivial changes by the adversaries? But it's worth looking into to see if it's feasible.
@tariqk @CobaltVelvet Oh it should definitely include bias...it should reflect the biases of the administrator, that's the whole point...taking the legwork and arbitrary nature of it all and making it consistent...
@CobaltVelvet @orielle don't disagree, should have been clearer: by “implicit bias” I mean, “bias I didn't know I had, and if I find out about it, I don't *want* to have.”
As for consistency... man. I wish I could be consistent long enough with my children so that they can pick up what I want them to pick up 😂😂😂
@orielle @tariqk Yes, I'm actually planning to build that in the future. :)