4 ms·
Systematic bad moderation can be tantamount to wokeness / cancel culture, especially if its manifestation in practice exhibits biases in the expected fashion.
by yanderekko 6y ago
Systematic bad moderation can be tantamount to wokeness / cancel culture, especially if its manifestation in practice exhibits biases in the expected fashion.
This is actually a major problem with arguments that try to dismiss the need to have ideological diversity in these institutions - cancel culture can arise from genuine good faith moderation if those moderators also happen to just have extremely biased ideological priors, eg. "we need to fight back" is taken as a literal threat of violence when said by the outgroup but not by the ingroup.
In this case, I think we need to be vigilant against NLP models meant to flag content that may not explicitly build in ideological biases but incorporate feedback loops that will reinforce them - eg. they're meant to detect posts that will likely be flagged by users / human moderators, but the baseline flag rates are ideologically-biased.
Unsurprisingly, the woke are perfectly able to understand these concerns when it comes to concerns about say AI coming to exhibit latent racial biases. But when the same mechanisms may cause ideological biases there's a telling lack of concern.
- chadash 6y ago> Systematic bad moderation can be tantamount to wokeness / cancel culture Maybe, but this is a bad example of it. Flagging posts that encourage suicide isn't a left vs. right issue... it seems pretty bipartisan to me. The fact that the AI made a mistake on a single example here isn't even indicative of bad AI... this could happen (and did happen in this case on appeal) with human moderators too.
- yanderekko 6y agoYeah, I do agree that in this instance you're probably correct. That said, it also isn't clear to me what in the tweet is easily mistaken for advocating self-harm - obviously keywords like "dying", "assisted suicide" and "failure" are probably playing a part but are not jointly sufficient. It could be that there's additional contextual information that could be seen as ideologically-valenced and that contributed to this moderation action, but we'll never really know. For example, maybe the tweet caused a lot of harsh backlash for ideological reasons and that makes it more likely for Twitter to action a post for any reason, and the model is just making a softmax prediction of what that reason is. That's something that we should find discomforting.