3 ms·
Yes, I'm convinced at some point we're going to figure out an algorithm to solve content moderation with some version of crowdsourcing like this based on reputa
by crazygringo 6y ago
Yes, I'm convinced at some point we're going to figure out an algorithm to solve content moderation with some version of crowdsourcing like this based on reputation, though I'd prefer a system based on building up trustworthiness through one's actions (consistently flagging similarly to already-trustworthy people).
But the challenge is still the same one you describe -- what do you do with competing "groups" or subcommunities that flag radically different things. What do you do when both supporters of each side of a civil war in a country consider the other side's social media posts to be misinformation and flags them? Or even just in a polarized political climate?
I still think (hope) there would have to be some kind of behavioral signal that could be used to handle this -- such as identifying users who are "broadly" trustworthy across a range of topics/contexts and rely primarily on their judgments, while identifying "rings" or communities that are internally consistent but not broadly representative, and so discount that "false trustworthiness".
But that means a quite sophisticated algorithm able to identify these rings/clusters and the probability that a given piece of content belongs to one, and I'm not aware of any algorithm anyone's come up with for that yet. (There are sites like HN which successfully detect small voting rings, but that's a far simpler task.)
- foerbert 6y agoI wonder if you could try to address this by limiting who can flag a given post. Even just doing it very naively and choosing, say, a fifth of your users for each post and only giving them the option to flag it might help significantly. It would probably make it more difficult to motivate the members of these problematic groups to actually coordinate if the average expected result was the inability to participate. And you could do it in more sophisticated ways too, and form flag-capable subsets of users for each post based on estimates about their similarity, as well as any other metrics you come up with - such as selecting more "trustworthy" users more often. This would help gather a range of dissimilar opinions. If lots of dissimilar users are flagging some content, that seems like it should be a strong signal.
- michaelmrose 6y agoPresumably instead of global moderation you could have pluggable meta moderation where you pick the moderators so you can have fun stuff like Moderator A whom you follow banned Bob therefore you can't see his posts or his comments on your posts but I don't follow A I follow B who believes Bob is a paragon of righteousness and so I see Bobs words everywhere and we all in effect have an even more fragmented view of the world than we have today with conversations that are widely divergent even within the same social media group or thread.
- jonnycomputer 6y agoSounds like a network analysis problem to tackle.
- ArneBab 6y agoIn the Freenet Project we’re solving content moderation with propagating trust — and visibility slowly increasing with social interaction, which is free for honest users (who just want to communicate) but has a cost for spammers and disrupters. The main tool is to drop the idea of a global definition of trustworthiness: You have seed IDs to find people who supply CAPTCHAs, but otherwise all trust is computed locally. If you want to try this, there are two steps: - First the current state in Freenet: https://github.com/xor-freenet/plugin-WebOfTrust/blob/master/developer-documentation/core-developers-manual/OadSFfF-version1.2-non-print-edition.pdf https://github.com/xor-freenet/plugin-WebOfTrust/blob/master... - Then the optimizations needed so this scales to arbitrary size: https://www.draketo.de/english/freenet/deterministic-load-decentralized-spam-filter https://www.draketo.de/english/freenet/deterministic-load-de... Here’s some data if you want to test algorithms: https://figshare.com/articles/dataset/The_Freenet_social_trust_graph_extracted_from_the_Web_of_Trust/4725664 https://figshare.com/articles/dataset/The_Freenet_social_tru... And some starting code of a more generic prototype for faster testing: https://hg.sr.ht/~arnebab/wispwot/ https://hg.sr.ht/~arnebab/wispwot/ (if you prefer a slightly larger writeup that might expand as time moves: https://www.draketo.de/software/decentralized-moderation.html https://www.draketo.de/software/decentralized-moderation.htm... )
- crazygringo 6y agoThanks for that, I'll take a look!