0:04 uh you know we built community notes because we wanted to build a better
0:07 informed world >> and uh as it scales to more parts of the internet that
0:12 means more people have access to accurate information. >> Great.
0:15 So let's look at a community note.
0:17 Um that's a note.
0:18 So what is it? >> Yeah.
0:20 So this is a real example of a community note we're looking at.
0:23 So basically here the post on the left is about Iran and it's saying
0:26 the USS Lincoln has been damaged and there's casualties
0:30 but actually the image is AI generated.
0:32 So this thing on the right here that says readers added context they thought
0:35 people might want to know.
0:36 That's a community note.
0:37 And what it's doing there is it's actually giving a lot of specific details
0:41 about what's wrong in the image.
0:43 And and it turns out that that level of detail that it goes into
0:46 is a big reason why people on both sides of the political spectrum
0:49 actually trust community notes more than a generic misinfo warning.
0:52 Uh the way these get here in the first place is they were actually
0:55 written by a regular user, a community notes contributor
0:58 and before they show on the platform to everyone and attach on the post,
1:01 they are rated helpful by
1:03 uh people from different perspectives.
1:05 So they're not shown unless that happens.
1:07 Uh another quick thing to call out is actually a lot of the best
1:09 notes are not just fact checks.
1:11 Uh they can add context
1:13 to uh posts that are correct but
1:15 uh otherwise >> Okay, so it's a context engine for news.
1:20 Uh but is it also for official accounts or ads or any kind of post? >> Yeah.
1:24 So a really important principle
1:26 of of the program is that all posts are eligible.
1:30 That means posts from heads of state,
1:32 posts from our company can get noted.
1:34 As Elon likes to point out, his posts get noted.
1:37 Um it regularly identifies AI generated imagery.
1:41 It's been a ton of that recently
1:43 with the Iran conflict.
1:45 It's detected deep fake audio of world leaders.
1:48 Uh it covers lighter subjects like entertainment, fashion, etc.
1:53 Um and as you can see up here, we've even had multiple notes on
1:56 both recent White House administrations.
1:59 Um and at least in one case, the White House actually took down the
2:02 post, issued an updated statement,
2:05 and you can imagine
2:07 like there was a person, a random person on the internet wrote that note.
2:10 You know, this isn't like a famous person.
2:12 They went out there, saw the White House said something wrong, typed in this
2:15 note, put it up there, and then suddenly,
2:17 you know, leaders of the free world changed their public statement.
2:21 That's like a superpower for people.
2:23 So, you can see why they're motivated to >> Yeah.
2:26 Well, so a a teenager I heard uh calls that a retraction, and it
2:33 really is a superpower,
2:34 but what is um the mechanism?
2:37 Uh can you take us back 10 years ago
2:40 before this superpower gets invented and distributed?
2:44 What caused the invention?
2:46 >> Yeah, I mean the origin for me goes back to 2016.
2:50 Um I was just a Twitter user then.
2:52 Um I was following the 2016 election.
2:55 Uh there were three televised
2:56 debates that year, but every day there was a debate on Twitter.
3:00 So that's where I was following.
3:01 That's where the world was following.
3:02 And I just, you know, I I remember getting a lot of good information,
3:05 but it was also hard to tell what was true.
3:08 >> And I was thinking just sitting on the outside thinking like, how is the
3:12 world going to solve this problem?
3:13 Like damn, like how are we going to do this in a way
3:15 that works and that people feel fair in a in a you know amidst polarization.
3:21 Um so then fast forward 3 years, I was working at Twitter at that
3:25 point and the industry had tried a lot of stuff at by then.
3:29 Um Facebook had built a huge factecking program.
3:32 uh Twitter was working with fact checkers and we also had internal
3:36 teams that would try to review posts
3:38 and you know decide whether they were were or were not misleading
3:42 and there were a bunch of issues with it.
3:44 It was just very clear these solutions were were not solving the problem.
3:48 There were issues with speed.
3:49 So typical fact checks just to put it in perspective
3:52 were often coming back in 2 to 4 days
3:54 which is like infinity and internet time.
3:57 >> Um scale was an issue.
3:59 Uh typically people could review like I don't know 10 order of 10 posts
4:04 or topics a day.
4:05 Um and even if you could solve those trust was the fundamental issue.
4:09 There were just a lot of people who
4:11 did not want or trust tech companies to be deciding what
4:15 was or was not accurate.
4:16 And so um I was managing a team at this time handed that off
4:20 and just went to go prototype crazy new ideas.
4:23 One of which became community >> Okay.
4:26 So the crazy idea just to play back is to think from the bottom up.
4:30 Um asking people to trust
4:33 uh random strangers uh on the internet
4:35 and amid a very high ppm
4:38 polarization per minute environment
4:40 as you just alluded to
4:42 uh why would people trust random strangers?
4:44 >> Yeah, it's a really good question and it's one we got all the time
4:47 getting started but uh the reality is people do trust community notes on both
4:51 sides of the polit political spectrum
4:53 and I think there's a couple big reasons why.
4:55 One is the process behind it.
4:57 So it's totally open, transparent, verifiable.
5:00 You can actually, and this is pretty wild in the world of social media.
5:03 You can actually download the real algorithm code that runs in production.
5:06 You can download the real data,
5:08 the community notes and ratings, run the code on the data to verify that
5:11 there's no funny business that we're doing on our end.
5:13 Like there's no override button.
5:15 Um, so it's really the,
5:16 you know, by the people.
5:18 Uh, I think secondly,
5:19 the notes are just really good.
5:20 So they speak for themselves.
5:22 They tend to be really accurate.
5:23 And the the main reason behind that is I think the the principle behind
5:27 the algorithm that doesn't
5:29 ingest any sort of external authority.
5:31 It actually decides what nodes to show by looking at
5:34 uh agreement from people who have disagreed in the past.
5:37 And sometimes we call that surprising agreement or
5:40 >> Okay, it sounds like time for a visualization.
5:43 Um so we have like two wings
5:46 uh and people post notes and then what happens? >> Yeah. Okay.
5:50 So just to orient you, what we're looking at here is every dot is
5:53 a community note and the the y-axis
5:56 there is how helpful
5:57 our algorithm thinks the note is and the x-axis
6:00 is the point of view of the note.
6:02 So you can see on the left and the right the red and blue
6:04 those are polarizing notes.
6:06 So only really found helpful by people on one side of the other and
6:08 it's super important that those don't actually show on the platform.
6:12 Uh actually the only notes that show to everyone are the ones in the
6:15 green oval at the top
6:17 and uh those are the ones that are found helpful by people who typically disagree.
6:20 And there's actually a bit of a community moderation element here too or if
6:23 you write too many of the notes on the bottom
6:25 you know you can actually lose your privilege for writing community notes.
6:28 And one thing that's really cool about this algorithm
6:30 uh if you compare it to
6:32 something like a more naive upvote down vote system like a majority rules type
6:36 of thing um you know something like that would just end up showing really biased notes.
6:40 And for here, our algorithm actually
6:42 takes advantage of partisansship and polarization.
6:44 So for any community note on a polarizing
6:46 topic, basically there's always going to be someone out there who's really predisposed
6:50 to disagree with that note.
6:52 Uh so you know, before they're going to rate it helpful, they're going to
6:55 go, you know, fact check it from every angle possible.
6:58 They're going to like really really check the sources in a lot of detail.
7:01 And and as a result, the notes that are actually found helpful in this
7:04 way uh tend to be really accurate, tend to use primary sources and
7:08 um tend to be, you know, pretty neutral in their language.
7:11 >> Okay, so people on both sides
7:15 um after turning polarization
7:17 into essentially fuel, right?
7:18 Geothermal energy uh uplift
7:21 something and both sides are happy.
7:23 But the person getting noted may not be very happy.
7:26 So say a a head of state
7:28 um gets noted uh and the head of state happened to have uh the
7:32 phone number of your CEO
7:33 uh and just calls
7:36 Elon and say take it down
7:38 uh by tomorrow morning.
7:40 Uh what would he say? >> Yeah.
7:42 So uh those you know emails like that calls or whatever they do come in.
7:47 Um fortunately the answer is really simple.
7:50 We have no override button.
7:51 Oh, >> so if uh if you're not happy with a note, you need to
7:54 take it up with the people.
7:55 Um and this was this was like kind of a crazy idea when we started.
7:58 You know, we went into a room full of like trusted safety people
8:01 and we're like, "Hey, so
8:03 the notes that show are going to be the ones that the people decide
8:06 and we can't take it down. >> There's no veto. >> There's no veto." Wow.
8:09 >> And you know, they're like, "What?
8:11 Like are you serious?
8:13 What if there's a bad note?"
8:15 Um but uh I think you know the point was if if it's the
8:19 tech company's opinion why is anyone going to trust it?
8:21 It needs to be the people's opinion.
8:23 And so we stuck to that principle.
8:24 Everyone got behind it and uh yeah we we have no we have no
8:28 way of changing the status of a note. >> Which is wonderful. >> Okay. It is wonderful.
8:33 >> And so what happens to the post after it gets noted. >> Uh yeah. Okay.
8:37 So what we're looking at here is on the left this is
8:41 a a typical representative post.
8:42 This is actually a real post and we're looking at how many people have
8:45 seen it over time.
8:46 So the the y-axis there is views and the x-axis is the time since post creation.
8:51 And you can just see this thing is going super viral at the start
8:54 uh all the way up until
8:56 it uh gets noted.
8:57 So that's the the green dot with the the yellow dotted line there.
9:00 Basically after that point it
9:02 um you know totally flattens out, gets almost no more views.
9:05 And the kind of crazy thing about this is
9:07 it's actually not getting downranked
9:09 by our for you algorithm. the post.
9:11 Uh this is actually just because
9:13 of what we call organic user behavior where basically people are
9:17 realizing now that the post is incorrect because the notes on it.
9:19 So they're just liking it less and reposting it less.
9:22 So I think this is really cool and one thing that I also love
9:25 is because our data is totally open.
9:27 Actually a lot of researchers from around the world have looked into this and
9:30 found the same thing.
9:31 So people from Stanford,
9:32 MIT, Udub um and Paris and Luxembourg
9:36 uh have have all actually
9:37 found a very similar thing that
9:39 reposts will drop by about 50%
9:42 or 2x after notes applied.
9:44 And this is this is you know this is really big in the scale
9:46 of social media like
9:48 uh you know 1 or 5%
9:50 when would be you know pretty big in and the scale of typical AB tests.
9:54 So one thing that I think is really heartening about this is that we
9:57 know from this and some other studies
10:00 that actually you know people are not just entrenched in their beliefs when a
10:04 note is applied to a post they'll actually agree with the core claims in
10:07 the post less and I think that's really cool
10:10 and um you know I I I guess there's a little bit of a
10:13 mixed blessing here though because
10:15 uh actually post authors will will also be more likely to delete their post
10:19 after they get noted.
10:20 So, in that way, the best notes actually get seen very infrequently.
10:24 Um, so I'm torn about that because just for me personally, you know, I
10:27 think not everyone agrees on this, but for me personally, I'd rather see a
10:30 post in a note than neither at all just because that's probably not the
10:33 only time in the world where you're ever going to see that particular wrong claim.
10:37 So maybe you'll see it off X somewhere in another post.
10:39 And just for me, seeing a lot of notes has kind of increased the
10:43 skepticism that I have when when reading things. >> Okay.
10:46 They serve as inoculation >> Yeah.
10:49 I think you know one other thing to mention about this graph
10:52 is it's kind of a big deal that this happens
10:56 people often assume the world is very polarized certainly feels very polarized
11:00 but people here are just they're just making a choice where they see a
11:03 post they see a correction they're like yeah things's wrong just not going to
11:06 share it and that's
11:08 happening across the political spectrum
11:10 um and we've seen that
11:12 pattern again and again when we first were designing the product
11:15 we did interviews with hundreds of people left and right and it was really
11:19 obvious that that most people just want to know what's going on in the world.
11:22 Um they are they know they're consuming incorrect stuff.
11:25 They just want to sift through it.
11:26 And this is just
11:28 showing that in action.
11:29 When given information, they're going to try to make a good decision.
11:32 And so I think people often assume like, man, it must be tough to
11:35 work on in this space of misleading information.
11:37 It must be, you know, must get
11:38 sad all the time, whatever.
11:40 It's actually like I feel very optimistic
11:42 working on it because we see there's quite a lot of agreement.
11:45 People are actually quite reasonable. So >> Wow.
11:48 Okay, so the ppm is going lower.
11:50 >> I hope I see it seems like it's lower than it might feel. >> Amazing.
11:55 So, let me now push
11:56 uh on a more cynical take.
11:58 Uh anyone who spend five minutes on the internet is probably thinking now there's
12:03 going to be a way to game this.
12:04 Maybe many ways to game this.
12:06 Um and just one example,
12:08 um I co-wrote a paper called malicious AI swarm.
12:11 Uh it talks about
12:13 um one person farming like 5,000 agents.
12:17 um the machine kind
12:18 uh and then they are some coded left some coded right they behave completely
12:23 normally they contribute even to committee notes
12:26 and uh just when the
12:28 controversial issue or election happens
12:30 then they manufacture a surprising
12:33 agreement and just note something that is actually true
12:36 uh how do you deal with
12:37 >> yeah manipulation's a a real thing I mean and people are always trying to
12:41 game social media algorithms and community notes is no exception
12:44 so I think one thing to call out is that surpris the agreement mechanism
12:48 does provide a bit of a defense against the a more naive attack than
12:51 the one you described.
12:52 Just there's a lot of people with the same view all piling on trying
12:54 to get a an incorrect note showing.
12:56 You know, that's not going to work.
12:57 But, uh, for a more sophisticated attack like the one you describe,
13:01 uh, I think we do have a lot of defenses in place.
13:03 So, just to name a few, we do things like requiring
13:05 a verified phone number from a trusted carrier just to increase the probability that
13:09 we're dealing with real humans.
13:11 uh we look for raiders who have rated things really similarly in the past
13:15 and uh actually we we we might treat them as the same uh person
13:19 just to to limit the influence of of really similar behavior.
13:22 Uh, another thing is we can look at
13:24 random samples of raiders
13:26 uh, and and if they're rating things very differently than self- selected, possibly malicious
13:31 raiders, then that's a very important signal.
13:33 And we have other things too like there's raider reputation to deal with lowquality people.
13:37 But I I think another key thing to call out is even even with
13:40 all the all these defenses, you know, committed notes
13:43 uh are incorrect sometimes.
13:44 Now, because it is really rare, we actually get this self-correcting
13:47 property where the uh the incorrect notes attract a lot of attention
13:52 and they'll they'll draw a lot of raiders to to go quickly rate them
13:55 not helpful and then they'll stop showing.
13:56 Uh and I think that self-correcting
13:58 property is super important also
14:00 just uh you know in a breaking news situation, right?
14:03 Something that was true a few hours ago may not be anymore.
14:05 So, it's it's great that notes are not set in stone. >> Okay.
14:09 So, notes being wrong is noteworthy.
14:13 uh and so people recursively improve.
14:16 Uh indeed I've uh seen it happening
14:18 on uh X for quite a while uh because I was a longtime contributor.
14:22 It just feels like magic.
14:24 It's like Wikipedia or Groipedia
14:27 uh when many people
14:29 swarm into some controversy.
14:31 It just gets really nice.
14:33 >> Uh but what about the other situation?
14:35 uh in a niche topic just developing
14:37 fast there just not enough attention to bootstrap
14:41 the initial surprising agreement.
14:43 So I've also seen like 5 hours 10 hours go by
14:46 without any consensus at all and so how are you going to tackle the
14:50 speed problem because as you pointed out if it comes next day it's already gone. >> Yeah.
14:55 Oh, well, first just to level set on speed.
14:57 I think Keith already mentioned, you know, the the previous state-of-the-art
15:00 factchecking would often take on the order of days and community notices
15:04 usually on more on the order of hours.
15:06 So, it is already quite a bit faster.
15:08 Nodes can appear as often as
15:10 about 20 minutes on a brand new post.
15:13 Uh, but they can actually appear instantly if there's already another node out there
15:16 that's uh matching on a URL or image or video.
15:20 Uh I think on top of that, one thing that people really really like
15:23 is if if someone actually sees a post
15:26 uh and engages with it before a community note is appearing,
15:29 we'll actually send them a push notification later so they get the correction as
15:32 soon as the note comes out.
15:33 U now even with all that, I think, you know, it's super important for
15:38 us to keep making community notes faster.
15:40 People want instant context and rightfully so.
15:42 So uh to that end, what we've done last year is we've actually opened
15:46 up an open API for AI contributors.
15:49 Uh, and this is a little bit of a crazy thing in the totally
15:52 open spirit of community notes.
15:54 Just like a regular person can be writing notes,
15:57 we let regular people write their own AI note writers and submit notes to our system.
16:01 Uh, and and what we've seen so far is
16:04 it's actually working really well.
16:05 The the notes are are really fast and and they're quite good, but
16:09 you know, definitely because it's AI, they're they're wrong some of the time.
16:12 Uh so so the way we we treat this now that's been working well
16:15 is we still have that human layer where where humans
16:18 write the notes in the same way as any other human author note.
16:22 Um and uh what we're working towards now is a way for AI and
16:25 humans to collaborate uh more effectively to co-write better notes
16:29 >> So humans are not just down voting or upvoting
16:32 but working with AI models. >> Yeah.
16:34 The idea is they can
16:35 like can we have humans and AI co-write these things together, co-create them together
16:39 and does that allow us to do this at a much
16:42 faster speed and and larger scale.
16:44 Um so yeah, what you're seeing on the screen here
16:47 is an example of what that looks like.
16:49 Um if there this is this is new.
16:51 Um if there's demand for a note, like people are requesting a note on
16:55 a post, AI will take a first shot at it.
16:58 Humans can write to you, but AI will take a shot.
17:00 Um and in this case, this is based on a real example.
17:03 the AI thought this video was from 2017
17:06 and it turns out it wasn't.
17:08 Um, so humans go in there and they're like, "Hey, actually like this is
17:11 from 2022 in Ukraine
17:13 and a bunch of people they they rate it and they make these suggested improvements."
17:17 They can also leave
17:18 uh suggestions on style or tone.
17:21 So they can say like, "Hey, I think this source is biased or I
17:23 think you should use a primary source.
17:24 It's going to be more trustworthy."
17:26 And then the AI takes that,
17:28 regenerates a note, and you know, usually gets it right.
17:32 >> And what's cool about this is first you get a better note on this
17:35 post people care about.
17:37 But two, all of those corrections,
17:39 all the suggestions are
17:41 training data that you can feed back into the AI.
17:43 So you can make it less likely to make that mistake again.
17:46 You can make it better at researching in the first place and also you
17:49 can make it more neutral, less biased.
17:51 Um, so all these human suggestions
17:53 both they they make better notes and they make better AI.
17:56 Okay, so just to
17:58 play it back, this is not
18:00 Grock helping humans to such a degree that it takes over all the judgment calls.
18:06 It is basically human teaching AI collaborative
18:10 learning so that the
18:12 translation between communities like
18:15 climate justice on one side, biblical
18:17 creation care on the other.
18:19 uh the AI model learns how to translate
18:22 and then and then what they become
18:24 better at this kind of translation.
18:26 Is this a new way to reward AI models?
18:29 How does it >> Yeah. Yeah.
18:30 There's this thing that we sometimes call reinforcement
18:32 learning from community feedback as opposed to just reinforcement
18:35 learning from human feedback which maybe would use like potentially a smaller bias set of non-representative people.
18:41 uh and and basically in the case of community notes
18:43 uh what it would look like is directly training the model to be writing
18:46 notes that would be maximally likely to be found helpful by
18:50 a simulated you know set of raiders who've typically disagreed in the past. >> Okay.
18:55 Well, that's that's really nice. >> Yeah, it's cool. Yeah. >> Yeah.
18:59 So still um someday I just open
19:02 X um and I just see
19:05 um peak slop uh like
19:09 the the marginal cost
19:11 uh for generating synthetic
19:13 media um even synthetic
19:15 intimacy uh is now falling so fast
19:18 and so it just sometimes
19:20 I feel uh personally
19:22 that whatever the corrective mechanism we invent
19:25 is not going to be fast enough
19:27 uh for this kind of pixelop situation.
19:31 So why should anyone here believe that computer nodes and collaborative nodes will evolve
19:36 to meet the demand? Yeah.
19:39 So definitely in the last 6 weeks or so with the Iran conflict,
19:43 we've seen the biggest
19:45 surge in in synthetic media that I've seen at least in the the
19:49 kind of misleading info space.
19:52 Um and I I will say like we're on the frontier here.
19:54 So uh this is the highest scale highest speed solution that exists.
19:59 Um these are new problems.
20:01 >> So you know we don't know what's going to work.
20:04 Can't guarantee the problem will be solved.
20:06 But I think there's a bunch of reasons to be optimistic
20:09 um for that problem like you know synthetic media surge.
20:13 We can both scale up the corrections
20:15 and we can both change the the fundamental
20:18 incentives and dynamics of the system.
20:20 So in terms of scaling corrections
20:22 uh we talked about AI
20:24 just like to put some numbers on that in the last 4 months alone
20:27 we've doubled the number of notes that are showing on X.
20:30 So this is like that's like not trivial
20:32 for a scale service to 2x in 4 months.
20:35 >> I think there's clearly headroom on that.
20:36 I don't is it 10x 100x
20:38 I don't know but there's clearly headroom to grow.
20:41 Um the other thing on the
20:42 incentive side, one of the reason
20:45 people post these things
20:47 is they can make money off of it through create uh creator revenue sharing programs.
20:52 And so we've recently put into place some changes to the policies there
20:56 where if you if your post is noted, you can't make money off it.
21:00 Also, if you post AI generated
21:03 uh footage of a war or conflict and you do not clearly call it
21:06 out, >> you are suspended from the revenue sharing program for three months.
21:10 If you do it again, you're suspended forever.
21:12 >> And so that's like kind of a big deal.
21:14 Those will will shape the underlying motivations people have. >> It's huge. >> Okay.
21:21 Um Chris, our curator,
21:23 uh when we talked earlier,
21:25 um asked this, I think very insightful
21:27 question because we've been talking about defense, right?
21:30 Defending against all those manipulations
21:32 and engagement through enragement
21:34 uh and so on.
21:35 Uh but is there a future
21:37 uh in which social media instead of pitting humans against one another
21:42 puts people and connects them uh with each other
21:45 elevating the voice that bridge
21:47 and you're like we have just a
21:49 >> Yes, we are building this. Yeah.
21:51 >> Um this is the this is the this is an awesome future.
21:53 So we have a pilot running.
21:56 Uh this is what you're seeing on the screen is what this looks like.
21:59 The idea is so in community notes we find the kind of like corrections
22:03 or context that's helpful to people from different points of view.
22:06 What if we could find the ideas or opinions
22:09 that are liked by people from different points of view?
22:12 And uh when when it happens in the pilot program, you see what you're
22:16 seeing on the screen here.
22:18 um the post will just get a call out saying liked by people from different perspectives.
22:22 And we see this obviously
22:24 people were were very happy to see Delta
22:27 not allow Congress to skip the TSA
22:29 line until TSA was funded.
22:31 Um and we see this
22:33 >> Uh yeah, you're you're among millions of people who also feel this way.
22:37 Um and and we see this agreement across a lot of topics.
22:41 We see like things that you think of as controversial.
22:43 We see it across immigration,
22:45 across the economy, taxes, international conflicts, etc.
22:48 Um there really is a lot of agreements out there, not on everything, but
22:51 there's quite a bit of it.
22:52 And um the concept is like what if we can
22:56 identify that like we don't need to boost this to start like just like
22:59 show people when there's agreement on something.
23:03 First of all, I think they'll find it interesting. It's a curiosity.
23:06 >> Um second, it it might incentivize more of that.
23:09 like maybe people will try to speak more in a way
23:12 where they can find out find that agreement
23:15 and and get more get more momentum behind those ideas.
23:19 >> Yeah, that's a really good point.
23:20 I think just in the same way that community notes spread less even though
23:23 there's no uh or community notes cause post to spread less even though there's
23:27 no downranking in the algorithm.
23:29 I think you'll probably see something analogous here where there's just a positive second
23:32 order effect from making that common ground common knowledge.
23:36 So it's a common knowledge engine
23:39 uh that turns polarization
23:41 into what we can all live with.
23:44 This is truly visionary
23:47 >> and think about it because this thing is open source.
23:50 It is open data.
23:51 So it means that not just X
23:54 uh but rather blue sky
23:56 true social everybody uh can just plug in that stream
24:00 uh and so that AI can learn from that and then connect the communities back together.
24:05 So what if we apply this
24:07 engine beyond social media?
24:10 Can you paint a picture of how that would look like? >> Yeah.
24:13 So where I mean where my head always goes is imagine just for like
24:16 one session of Congress,
24:18 everyone just focused on delivering where there was
24:21 >> you know, whether it's immigration, taxes, whatever.
24:24 I think people would be stoked. And uh yeah.
24:30 >> There's a lot of agreement on these topics.
24:32 If all we did was pursue the areas for agreement, I think people would
24:35 be pretty happy with the direction the world was going.
24:37 And so, um, you know, my hope is with programs like this,
24:41 uh, if we can
24:42 identify common ground at internet scale, it'll make it a lot easier
24:46 to create a future that humanity likes.
24:49 And so hopefully we can help with that.
24:54 And with that, uh, Jay, Keith,
24:57 thank you for being our best builders
24:59 and showing us that a pro-social
25:02 media future, uh, is not in some sci-fi is already here. Thank you. >> Thank you. >> Thank you, Audrey.
25:12 >> I'm Audrey Tong and I'm a guest curator
25:15 along with the fantastic
25:16 Divia Sedarth at TED 2026.
25:19 As guest curators, we get to bring people who are doing incredible
25:22 work into the TED stage,
25:24 help them find ways to share that work with the world,
25:27 and be able to create a dialogue between what we think are some of
25:30 the best ideas out there and solving the problems we care about the most,
25:32 AI, democracy, these big questions,
25:35 and the TED audience and really the wider world.
25:37 And we chose the interview format
25:40 to bring Keys and Jay in
25:42 because we really feel
25:45 that the 18 minutes talk format
25:47 as good as it is
25:49 not doing the full justice
25:51 of their job which is training
25:53 an AI to understand
25:55 the differences between say climate justice communities
25:59 and the biblical creation
26:00 care communities and the various different aspects
26:03 that this social translation
26:05 can do to our democra
26:07 Y so I try to push them like really hard
26:10 in every answer they give and they took it like a champion.
26:13 I think one of the great things about this talk is
26:16 you know a lot of it is about community notes which is a fundamentally
26:19 defensive approach right we understand that the world is full of lots of bad
26:22 information we try to prevent the bad stuff from spreading
26:25 but I love the ending which is on well what would it look like
26:28 if we flipped this and I hadn't thought about that as much before where
26:31 if we flipped this to say as much as we know the kinds of
26:34 corrections people agree on we could also figure out the kinds of information and
26:37 positive solutions people agree on and make that actually be the thing that people
26:41 are focused on online instead
26:44 all the other stuff that they tend to focus on online.
26:46 >> This ending talks about data is soil.
26:50 So that the understanding
26:52 between different communities tend together
26:55 this garden of AI agents
26:57 that grow with our communities
26:59 loyal to communities and not trying to extract anything but just to regenerate our deep understanding.
