Disagreeing with the decisions of moderators is orthogonal to comparing AI and Human moderation.
AI moderators are terrible at enforcing a complex set of practices and rules.
Human moderators are excellent at enforcing a complex set of practices and rules.
Both are subject to the whims of those defining the practices and rules, and their enforcement will always fall within scope of those definitions. If Twitch’s moderators are given full empowerment to set the rules they moderate by, then it is appropriate to blame them for what is perceived as faulty moderation decisions. Chances are, they do not have any such authority, and at best can “recommend” or “advise” while being compelled to silence by their overlords.
You are absolutely right that humans make biased decisions, no matter how much we work to correct for it; judge sentencing becomes more strict when they’re hungry or their local sports team lost, and referees are less likely to penalize the home team when in the home stadium with home fans yelling at them. But these biases can be documented, studied, and gradually accounted for in training and post-decision reviews. It’s not perfect, but it’s still great.
With AI moderation, we bake the biases into the dataset and training models, and then we throw complex rule making decisions at a nascent ML network that can’t discern the simplest of decisions that any human moderators could: “Is this erotic or non-erotic content?”. Setting aside the biases of any possible training set versus Rule 34, it is dead simple to outwit any AI in this regard, in ways that most human moderators would detect instantly. It’s not perfect, and in fact it’s pretty awful.
I agree that the rulemaking decisions of modern sites are often terrible, biased, and exclusionary of entire categories of human beings and behaviors. However, the errors made by human moderators trying to service the complexity assigned to them do not somehow excuse the staggering naivety and incompetence of AI moderators. The heavyweights in this Ml moderation industry — Google, Amazon, Facebook — have failed to deliver effective ML moderation after billions of dollars and cumulative decades of investment, and paper over their failure by using user flags and human review teams to disguise the sharp edges of their solutions.