Community Notes: Scaling Public Epistemics by Jay Baxter | Devcon SEA

[Music] [Music] baby at home hopefully this can be just as good as if I was there um uh you probably already are familiar with the community notes UI but just in case you're not um here's a note attached to a post on X um this note was written by a contributor a basically regular user and before it was shown was rated as helpful by other contributors um who tended to disagree in the past um this lightning icon means that the note actually appeared within the first uh hour of the post being created which is happening more more since the system's been speeding up uh recently another way notes can appear really quickly on posts is via media matching so you can write notes on um on images or videos that will automatically apply to all their posts with those uh uh that same matching media um because the data is public uh external researchers have analyzed a ton and I'm just going to share some highlights of what they found so one group uh found that notes are are 80 98 or 99% accurate about covid vaccines according to them uh and just want to F that although notes are already pretty rarely inaccurate uh one great property is that uh inaccurate notes attract additional Raiders and as a result the system can quite often be self-c correcting um which is pretty cool um another group ran some a survey where they compared some uis um one that's kind of like Community notes versus um one where they had a Community flag or a traditional fact checker flag and what they found is um that the community note is found uh bipartisanly by surve par participants to be more trustworthy than the others and then also um and Survey participants rate posts as less accurate when they're displayed with a note versus without um which is pretty cool we uh do not algorithmically rerank posts based on whether they have notes or not for instance in the 4u page but there is is an organic uh change and engagement Behavior after post get notes so just uh we observed in 2022 an enab test that the um like and repost rate dropped by 30% um uh after a post got a note um and then in 2023 and 24 um external research groups found uh about a 50% drop in engagements after a post got uh noted so that's really big by the way if you've ever ran experiment on big social media platforms you're probably familiar that 1% can often be a really really big change and these are these are you know in the 30 30 and 50% range um so it's quite large uh posts are also much more likely to be deleted by their authors and then this is really hard to measure but there's just an incentive change uh authors don't want to get demonetized or have the stigma of a note so um in otally some users say they post differently um given that notes exists I think very relevant for this top is the idea of credible neutrality that we aspire to um similar to um some other crypto projects um I think um some things we do towards this end are number one contributors are regular users uh right they they do have to have a verified phone number and have signed up six months ago Etc but for instance they're not employees or something like that um separ we don't take in external ground truth labels either so instead of we use this clever bridging mechanism that'll talk more about uh and then uh we also are open source transparent reproducible and verifiable so for instance X never manually edits or changes the status of a note to take it down uh you can actually go to our website and download the code download the data and run the code on the data and verify that it matches exactly what was in production two days ago that two-day delay is just to prevent manipulation um but that's pretty cool to be quite uh so transparent and is very rare among some more companies um okay so the core mechanism I think that's interesting here is this bridging algorithm um which is what lets us not use external ground truth um so here's a note of one of our contributors explaining it to someone else uh doesn't work by majority rules uh it's not an up vote down vote system where we add it up we we actually look for uh ratings where there's agreement between people whove disagreed in the past um this gives us some built-in manipulation resistance that's pretty cool uh if in this little cartoon here um just to help you understand why this might happen uh is if you think of these circles as people and the colors are viewpoints uh if you get some off platform berrade of people from Discord or something who all decide to go rate a note a particular way uh you know that group alone can't determine what notes get shown um it's a pretty great property um uh another interesting thing is that the data set's quite large so it's it's sparse there's not a lot of users that have rated many of the same notes in the past so what we have to do is we build a model we uh use a matrix factorization approach so we construct The Matrix of users and and notes and the ratings on them and we uh train this model that's quite similar to uh a traditional recommender system you'll probably familiar with this but rather than personalizing what note we show to each user we actually learn this Global note score um uh which we call an intercept term here um which is the same for all users um and the way we train it is we add extra regularization to those terms which force ratings to be explained by the viewpoints whenever possible so if the if ratings are um uh occurring on a note in a totally Viewpoint consistent way then The Intercept term will remain neutral uh what this looks like in picture format is here's a scatter plot of notes um the x-axis is the Viewpoint of notes the y- axis is The Intercept score um most of these notes fall in the yellow circles and that's not um a problem with the algorithm often times these notes um just aren't great um uh the the green notes uh are the ones that are seen by users on X um everything else is just seen by contributors and then the red ones we actually penalize the author is for um so that's the core mechanism and how it works uh there's a ton of other mechanisms on top of that um which I don't fully have time to explain all of them but for instance uh we also do a similar Matrix factorization for tags so if we find bridging based consensus that a note is incorrect uh we don't show it even if it was found helpful for instance there's also needs your help tab uh this can and needs your help notifications so this can help uh number one address selection bias we can make sure there's enough Raiders with each Viewpoint that have rated a note to help get notes showing faster um we can detect or mitigate coordinated coordinated manipulation like those brigades um if the distribution from the needs or help tab is very different than the organic distribution um I think of particular interest to this group is probably similarities uh with prediction markets so you can think of each note as being kind of like a prediction Market that resolves to its final status uh whether it's helpful or not helpful and your R reputation goes up if your matches the outcome and down if it doesn't so the Rader reputation is kind of like your currency in the prediction Market um the uh the the market doesn't always resolve of course um as prediction markets also don't always um which is fine it's it's a vowed distilled mechanism um uh in that case and I guess maybe the biggest difference is that there is no bet sizes um so each user gets one rating Max um I think uh the incentives for reputation well first call out that some users just aren't very incentivized by and they're just going to uh rate in ways that uh they're going to try to impact what noes shown in the platform above all else but as far as users who are following the incentives number one for your ratings to count you have to have high enough reputation uh number two in order to earn the ability to write notes you have to have high enough raing impact uh and then uh in order to unlock some super superpowers like media notes if the build up your writing impact I think one pretty straightforward idea here could be just making prediction markets that resolv to whether there's Community notes on a post uh but I think uh you know vitalic has been blogging a lot recently about um Community notes and prediction Market mashups um and you know what's the what's the best way and I think many others are excited about this what's the best way to um build community notes that could say something interesting about the future um What mechanisms could work there um I think I just want to end the talk by saying things that I'm personally interested in I think uh those mechanisms um and as well as multi-dimensional bridging um more notes uh faster in more places like matching and llms and um you know just to uh I think bridging algorithms um Beyond notes like in the foru page for instance would be cool um or collaboratively writing wikis or or uh policies um you know no guarantees this is just my personal uh interest so no guarantees about the company um doing anything there but uh if any of this is exciting to you uh we're hiring a machine learning engineer um it's also linked to my bio and definitely especially because I didn't get a chance to come there in person please uh reach out and DM me with any questions or comments you have thank you so much thank you Jay