Honestly, content moderation is one of the things where I kind of see a legitimate application for AI. Like, I don't actually think the technology is good enough to do it effectively, but I also think that human workers are already incapable of doing it effectively, and at least you wouldn't be forcing a bunch of people to look at torture, rape, murder, and hate propaganda all day, so it seems like a net improvement.
While I see what you’re saying, in order to train the AI to spot that stuff, the method is to force a bunch of people to look at torture, rape, murder, and hate propaganda all day.
Oh yeah, yeah, I see the problem.
But it stays trained after you do it, right?
Ideally, yes, but then it hits the steady-state mathematical accuracy of 70% inherent to the linear algebra that powers all AI.
@esoteric-merit, you’re more familiar with this stuff, so I’m tagging you in. (Also you are both two of my favorite mutuals, now kiss interact)
Me and Quasi already interact a fair bit, actually. Yeah, there's an end problem where our current identification processes are all, when it comes down to it, complicated linear regression models. And no matter how good a regression model is, it has a theoretical maximum. Call it AI, or Neural Networks, or Machine Learning, it doesn't matter, they're all bound to the same limits. And that limit is low enough that . . . you would still be forcing people to watch torture & etc. all day. But it would *also* be bad enough to let tons of it slip through. So you get the worst of both worlds. The statistical tests that do provide better accuracy & precision, like, say, monitoring accounts that have posted it before, performing fuzzy hashes on known pieces of content and comparing those hashes to uploaded media, looking for certain words or tags popping up, are already being done in most places and never needed machine learning to enter the picture. There is an add'l problem in that once a machine would be in place, people would soon learn exactly what sneaks entirely past the machine and never gets flagged in the first place. With human moderators, you have to fool all of the (constantly updated) tag searching and account/IP monitoring, and a lot of stuff you can't actually get away from, and then a human might still see it, realize it's hiding from the system, and metaphorically pick up a log to reveal all the insects writhing beneath. Then they update all of their existing statistical tests, and another wave of stuff is removed without (much) human intervention anyways. . . . but once the machine is fooled, people can figure it out, even if only through superstition, and you've offloaded things onto the machine, and now it never self-corrects because you've removed the apparatus by which the system did so. So you need to review raw content *anyways*. So it can't even functionally act as a force-multiplier, because you already have a lot of force-multipliers, and the trouble is how quickly they're side-stepped. Which ML would be worse at. SO. THAT ALL SAID. You might still be seeing where it can be *part* of those force-multipliers, not trusted, but acts as one of the weights. And guess what? It's already been used like that for ages! A couple of decades now, to my knowledge. So I haven't been explaining why it can't be used, I've been explaining why it's not a magic bullet, and would still lead to our current situation, (and can't really meaningfully improve without an entire unforeseen paradigm shift in machine learning algorithms) . . . which I can state extremely confidently because it is in fact already part of our current situation.
Thank you, that was very informative!
