The model-based decision making is likely not capable enough. Specially not for the way that Reddit Inc. would likely use it - leaving it in charge of removing users and content assumed to be problematic, instead of flagging them for manual review.
I’m specially sceptic on the claim in the site that their Hive Moderation has “human-level accuracy”. Specially over time - as people are damn smart when it comes to circumventing automated moderation. Also let us not forget that the human accuracy varies quite a bit, and you definitively don’t want average accuracy, you want good accuracy.
The model-based decision making is likely not capable enough. Specially not for the way that Reddit Inc. would likely use it - leaving it in charge of removing users and content assumed to be problematic, instead of flagging them for manual review.
I’m specially sceptic on the claim in the site that their Hive Moderation has “human-level accuracy”. Specially over time - as people are damn smart when it comes to circumventing automated moderation. Also let us not forget that the human accuracy varies quite a bit, and you definitively don’t want average accuracy, you want good accuracy.
Regarding the talk about biases, from another comment: models are prone to amplify biases, not just reproduce them. As such the model doesn’t even need to be trained only in a certain cohort to be biased.