[00:00.830 --> 00:16.890]  And one of the things to really think about that has transpired over the last, close to a decade now, is the revolution of the social media ecosystem has gone through since really around 2016.
[00:17.170 --> 00:22.850]  Although, obviously, places like 4chan and other forums have been around for much longer.
[00:23.610 --> 00:29.710]  I'd like to compare it to the change that we saw in our information landscape.
[00:29.990 --> 00:40.610]  We moved from three main channels, so equate ABC, NBC, and CBS to Facebook, Twitter, slash X, and TikTok.
[00:41.770 --> 00:47.510]  These are places where most people were getting their news, their sports, their celebrity gossip.
[00:47.750 --> 00:51.490]  All of the kind of content all centralized in one place.
[00:53.150 --> 01:06.230]  And instead of going to one place to get all of your information, you're now picking and choosing from the incels.is, or you're going to read the latest and greatest stuff on True Social.
[01:06.610 --> 01:06.910]  Right?
[01:07.230 --> 01:19.170]  You've moved away from centralized information on the main platforms to a world where you're picking and choosing the kinds of content that you want to see specific to your interests.
[01:19.170 --> 01:26.150]  So, if you want to watch 24 hours of the crocodile hunter, you can do that.
[01:26.270 --> 01:26.490]  Right?
[01:26.530 --> 01:29.430]  You can do that on cable with the nature channel, I think.
[01:31.050 --> 01:45.050]  But instead, if you wanted to read white nationalist tracks or be immersed in extremist ideology, you can now spend all of your time in specific telegram channels and get that into your bloodstream all the time.
[01:47.070 --> 01:54.690]  And that change has been accelerating over the last, let's say, five years.
[01:55.150 --> 01:58.590]  And some of you may be familiar with some of these sites.
[01:59.570 --> 02:11.130]  You know, most of the time when we talk about the alternative social media ecosystem, many folks will say, oh, I never even knew MeWe existed or I didn't know what DLive was.
[02:13.190 --> 02:16.330]  There are so many places now where you can get content.
[02:16.490 --> 02:16.670]  Right?
[02:16.770 --> 02:22.670]  If you're a militia here, you can go to MyMilitia.com and hang out with other militia folks.
[02:22.910 --> 02:23.010]  Right?
[02:23.190 --> 02:32.590]  So it's both a factor of the community building that's happening in those places, as well as the echo chambers that's developing in terms of the belief systems that are there.
[02:34.670 --> 02:44.150]  And it's not just the fact that there are so many of the sites, it's the number of folks who are on those sites and getting all of their information from those sites.
[02:44.150 --> 02:59.090]  So if you look at Pew Research polls over the last two or three years, they've been really deep diving into the communities on Gab, on Truth Social, on some of the smaller platforms.
[02:59.610 --> 03:11.530]  What they've discovered is something on the order of four to five percent of American adults get all or almost all of their news information from those sites.
[03:11.530 --> 03:15.450]  Now, four to five percent may not seem like a large number.
[03:15.890 --> 03:33.330]  But when you start to extrapolate out how many people that actually is 10 to 15 million folks who are only getting their information from some of the poorest, I would say poorest in terms of their enforcement of their own terms of services, as well as doing any kind of content moderation.
[03:34.070 --> 03:42.790]  It's no wonder that we're getting more and more extreme polarization in our country in terms of the kinds of belief systems.
[03:43.070 --> 03:48.650]  And we just saw an event happen in Butler, Pennsylvania.
[03:49.370 --> 03:54.810]  Former President Trump was, it seems like, injured in some way.
[03:56.070 --> 03:56.870]  We don't know.
[03:56.970 --> 03:56.990]  We don't know.
[03:56.990 --> 04:03.850]  I'm getting a lot of questions from folks asking me if we saw anything in the lead up to that event on these platforms.
[04:04.410 --> 04:05.210]  I don't know.
[04:05.390 --> 04:07.470]  I haven't seen any evidence of that.
[04:07.470 --> 04:26.790]  But the amount of extreme acceptance of political violence, of basically legitimizing political violence, that's happening on these platforms, that's of concern.
[04:27.510 --> 04:30.310]  And it's happening not just on the far right.
[04:30.690 --> 04:34.970]  We are also seeing this, and more and more folks are talking about it.
[04:35.110 --> 04:45.930]  It's happening on far left, even on some of the more centrist, you know, places ideology-wise, where people are saying things like, bring it up.
[04:47.050 --> 04:52.550]  I'm ready for this civil war or event to happen.
[04:54.350 --> 04:57.670]  And yes, a lot of this is just shitposting.
[04:59.970 --> 05:09.510]  But when it starts to become a fever pitch, I think that's when people should really be concerned about how many of these people actually believe this.
[05:10.350 --> 05:14.610]  And I'm going to talk a little bit, you know, more on mass shootings.
[05:14.610 --> 05:24.710]  Here are some of the statistics in terms of how users who now have, at their fingertips, a content delivery network that's really, really hard to take down.
[05:27.910 --> 05:33.070]  More and more people are being exposed to extremist ideologies, to hateful content.
[05:33.390 --> 05:43.650]  Something our company helps our NGO clients, our human rights clients monitor, try to understand threats to their communities, threats to ethnic groups.
[05:43.650 --> 05:52.570]  Ever since October 7th, you know, things have exploded, both in terms of anti-Muslim and anti-Semitic content on these platforms.
[05:52.570 --> 06:04.350]  You know, and I would say, yes, there's a correlation between the rise in this type of content and the rise in actual real-world physical hate crimes and things like that.
[06:04.490 --> 06:08.590]  We can't draw a causatory, you know, arrow between those things.
[06:08.590 --> 06:26.790]  But it's not hard to imagine that being immersed 24-7 in hateful content is potentially going to impact your mental state in terms of willingness to, you know, conduct a attack against somebody on the street.
[06:27.650 --> 06:36.170]  And this is more and more, I think, a concern when it comes to, you know, especially on the anti-Semitic and anti-Muslim.
[06:38.590 --> 06:44.010]  Now, deliberately left off the names of many of these individuals on this slide.
[06:44.150 --> 06:56.410]  But I think one thing to highlight here is the pathway through which Peyton Gendron, the shooter at the Topps grocery store in Buffalo, New York, got to...
[06:59.640 --> 07:16.340]  He was first and self-reportedly exposed to video of the Christchurch shooting by Bretton Tarrant, who had live-streamed on Facebook Live before Facebook decided that it was probably not a good idea just to let people willy-nilly stream things, live-stream,
[07:16.560 --> 07:17.940]  on their platform.
[07:20.440 --> 07:27.100]  That video, which I've unfortunately had to watch more than a dozen times, just really sticks with you.
[07:27.100 --> 07:30.020]  It's a massacre in a mosque in Christchurch.
[07:31.020 --> 07:36.460]  And Gendron viewed a clip of that on 4chan, became curious about what that clip was about.
[07:37.700 --> 07:40.640]  He started to self-radicalize, went down the rabbit hole.
[07:41.720 --> 07:46.600]  Eventually, when he conducts his attack at the Topps grocery store, live-streams it as well.
[07:47.140 --> 07:49.720]  Clips of that are still readily available, right?
[07:49.760 --> 07:51.920]  It's really hard to take all this content down now.
[07:52.560 --> 07:55.000]  Even with all of the hashing techniques that are out there.
[07:55.640 --> 07:59.580]  And also tries to mimic Tarrant's manifesto.
[08:00.300 --> 08:03.900]  And if you look at the... basically the playbook now,
[08:06.920 --> 08:18.180]  you know, before they conduct their attack, they'll go on one of these smaller platforms, they'll post something, and a bunch of people will just download it, right?
[08:18.240 --> 08:19.160]  And it just lives forever.
[08:19.520 --> 08:27.820]  You can, right now, go on, you know, probably once you go through a couple layers of Google searches, you can find Bretton-Tarrant's manifesto.
[08:28.360 --> 08:30.760]  And someone out there is going to read that.
[08:31.080 --> 08:39.560]  And, you know, it's also what inspired Gendron and, you know, Andres Breivik inspiring Bretton-Tarrant, right?
[08:39.700 --> 08:46.080]  It's a cycle that is enabled by the fact that it's so hard to get this stuff off of the Internet.
[08:47.940 --> 08:54.360]  All of this is leading to a higher red environment than I think we've seen.
[08:54.360 --> 08:59.040]  You know, obviously the presidential election that's coming up is a big catalyst of this.
[09:00.620 --> 09:11.320]  But we have seen many of these platforms be used by nefarious actors for other purposes, you know, that are not related to political violence, right?
[09:11.520 --> 09:14.660]  There's child sexual exploitation happening on places like this.
[09:14.800 --> 09:16.540]  They understand that it's a major problem.
[09:17.240 --> 09:20.560]  They're still having problems internally trying to stamp this out.
[09:21.820 --> 09:27.880]  Gab, which was the host of, you know, some of these manifestos as well, right?
[09:28.620 --> 09:29.520]  The tree of life shooter.
[09:30.600 --> 09:37.900]  You know, again, many of the platforms that we're pulling in data from are not even enforcing their own terms of service.
[09:38.140 --> 09:42.140]  They're stated terms of service when it comes to what's allowable and not allowable.
[09:42.140 --> 09:44.980]  And I'm going to talk a little bit more about Telegram.
[09:45.140 --> 10:03.920]  Telegram's role in an Atomwaffen plot to conduct attacks against electrical substations in Baltimore to essentially cause an accelerationist race war happen faster in Baltimore.
[10:03.920 --> 10:16.740]  You know, but again, the platforms are enabling a lot of this very, very difficult, you know, content that should be removed.
[10:17.000 --> 10:24.580]  And I think that's a major part of why we need AI essentially to help sift through all this content, right?
[10:25.120 --> 10:29.300]  Our platform, we pull in something like 6 million posts every 24 hours.
[10:29.860 --> 10:47.360]  If you think about what a human is able to look through on a daily basis, never mind sleeping and getting, you know, some kind of mental health break for having to look through all this terrible stuff, something like maybe 600 posts a day is the maximum somebody can actually credibly look through.
[10:48.980 --> 10:51.760]  Six million is orders of magnitude more than that.
[10:51.960 --> 11:04.100]  And even if you have hundreds and hundreds of analysts looking at all this content, you really do need to have some, you know, artificial assistant to basically help you look through that content to make sense of what's going on, understand the truth.
[11:05.460 --> 11:20.740]  So, one way of thinking about how these threats are manifesting is to connect how virtual content can potentially lead to physical, real-world events.
[11:34.910 --> 11:45.030]  The kinds of cybersecurity threats that we're used to seeing as an industry, as a community, happen purely in the virtual realm, right?
[11:45.030 --> 12:02.070]  These are basically bits and bytes that are then leading to, you know, more bits and bytes being deployed, where a breach happens and that data is being hosted somewhere and everything happens at the virtual layer.
[12:03.430 --> 12:12.290]  Where social media comes in and user-generated content being a threat, that's really happening at that virtual to physical level, right?
[12:12.390 --> 12:19.390]  Where you see Malign organizing on the kinds of events such as what happened in Ottawa, Canada, right?
[12:19.690 --> 12:21.770]  And you got the trucker convoy there.
[12:22.610 --> 12:39.730]  One of the really interesting things with the trucker convoy was watching how that threat evolved over time and deciding to franchise essentially a trucker convoy in the United States whose first threat was against the Super Bowl in Los Angeles that year.
[12:40.210 --> 12:46.350]  Eventually, those folks made their way from Los Angeles all the way to Washington, D.C.
[12:46.710 --> 12:50.370]  and tried to drive around the Beltway a bunch of times and people just honked at them.
[12:50.650 --> 12:51.210]  Kind of ridiculous.
[12:52.670 --> 12:57.450]  And then there's things that happen in the physical world that lead to virtual world propaganda.
[12:57.450 --> 13:24.170]  So, thanks to the war in Ukraine and the kinds of false information, October 7th as well, and what's happening in Israel-Palestine, these are physical world events that are then leading to people essentially recrafting the digital ephemera from those events and posting them on virtual platforms to
[13:24.170 --> 13:29.490]  inform people, misinform people, to disinform people, right?
[13:29.690 --> 13:35.590]  That's, again, an example of how the physical world events then lead to virtual content.
[13:38.090 --> 13:43.410]  And I think we shouldn't underestimate the speed at which some of these things are happening.
[13:44.290 --> 13:47.770]  And I'll point to, for example, what happened to Okta, right?
[13:47.830 --> 13:48.610]  There was a big breach.
[13:50.270 --> 14:07.730]  Within, I think, hours of the breach happening, the group that conducted the breach was already going and advertising on Telegram saying, hey, you know, here's some evidence that I've got this breach data and I'm willing to sell this to the highest bidder,
[14:08.010 --> 14:08.110]  right?
[14:08.230 --> 14:14.450]  And there's tons and tons of user data, very sensitive data available there.
[14:14.910 --> 14:23.470]  In terms of reputational threats, we've seen companies go after, like, Etsy, Coca-Cola, McDonald's, Starbucks.
[14:24.130 --> 14:43.410]  And then, of course, there's the threats that are coming from places like Kiwi Farms and Incels going after just private citizens, people who are not used to or equipped to deal with the deluge of virtual threats that they're receiving and trying to figure out which of these folks is actually serious about causing me physical harm, right?
[14:46.490 --> 14:51.950]  And obviously, what's happened with Bud Light and Target over the last year or so.
[14:53.350 --> 14:59.330]  A case I'm going to draw your attention to here is what happened in Lewiston, Maine with Robert Card.
[14:59.330 --> 15:01.690]  Robert Card was a military veteran.
[15:01.890 --> 15:09.770]  He was an Army National Guardsman who conducted a mass shooting and killed 18 people.
[15:10.110 --> 15:19.330]  And there was an incredible amount of speculation about Card's mental state, about his motivation, about his whereabouts, right?
[15:19.410 --> 15:23.270]  Because he was not immediately apprehended after the shooting.
[15:23.270 --> 15:39.790]  And one of the things that occurred on places like True Social, on places like 4chan, were people trying to dox Robert Card and posting the information of the wrong Robert Card on those sites, right?
[15:39.790 --> 15:44.170]  So again, causing harm to somebody who was innocent of this specific crime.
[15:45.010 --> 15:53.110]  And also posting, you know, potential whereabouts of Robert Card and, you know, false information about his vehicle and things like that.
[15:53.110 --> 15:58.230]  And getting National Socialist Club, neo-Nazis, right?
[15:59.390 --> 16:04.650]  Posting on their channel, hey, we have information about Robert Card where his whereabouts are.
[16:04.750 --> 16:06.790]  Everybody come out, let's go get him, right?
[16:07.210 --> 16:12.470]  Again, threats to public safety happening, organizing happening on these platforms, right?
[16:12.710 --> 16:16.170]  Last known location, here's a very specific address, right?
[16:16.170 --> 16:18.150]  And what if there was a Subaru there?
[16:19.130 --> 16:21.970]  I mean, who's going to take that action into their own hands?
[16:22.530 --> 16:23.810]  Be responsible for that.
[16:23.970 --> 16:28.110]  That's the kinds of things that are enabled by these platforms.
[16:30.370 --> 16:42.150]  Adam Wofford, in their own White Nationalist Accelerationist channel, posting propaganda, posting content about how to take down electrical substations, right?
[16:42.150 --> 16:48.410]  This is part of... I don't... Some of you may be familiar with the term Rahoa, racial holy war, right?
[16:48.450 --> 16:58.810]  It's part of this White Nationalist ideology of trying to promote an accelerationism, make more likely and make faster a war between the races.
[16:58.830 --> 17:02.430]  They targeted Baltimore because it's heavily African-American population.
[17:03.190 --> 17:11.350]  And, you know, I think a number of these folks were just arraigned on this plot, right?
[17:11.350 --> 17:15.610]  And again, they were promoting this content on Telegram, on their channels.
[17:17.170 --> 17:19.350]  So, the evolving threat landscape.
[17:19.530 --> 17:24.350]  And I'm going to save the rest of my time for questions because I'm sure there are some.
[17:24.370 --> 17:26.850]  And I'd love to have a conversation about this.
[17:29.090 --> 17:41.730]  We're tracking something like 65 fairly large social networks that are out there that have emerged over the last, you know, decade or so that are not being monitored by, you know, most of the major platforms that are out there.
[17:43.270 --> 17:52.450]  That is an incredible burden on an analyst to try to even figure out, like, where is the newest site where people are posting stuff?
[17:54.310 --> 17:59.850]  We get asked all the time, hey, what new places are you seeing?
[17:59.850 --> 18:05.410]  Because we want to figure that out for ourselves and try to go and access that data, right?
[18:06.070 --> 18:09.090]  And it's not just text, right?
[18:09.150 --> 18:20.670]  Our system is able to use AI to categorize images for things like KKK-type content, anti-Semitic content, anti-LGBT content.
[18:22.130 --> 18:28.130]  And it's necessary to do that because sometimes the threats are coming in the form of a picture, right?
[18:28.390 --> 18:40.230]  So, an example, we had a client who one of their employees had been very vocal about an activist stance that they took.
[18:41.090 --> 18:49.670]  And 4chan decided, oh, yeah, we're going to, you know, teach this person a lesson, found their photo and posted their photo along with their personal information.
[18:50.530 --> 18:51.450]  But no name.
[18:51.910 --> 18:55.130]  So, if you're only looking at the person's name, you're not going to be able to see that, right?
[18:55.130 --> 19:06.850]  You have to be able to scan the photo and understand, you know, run facial recognition on it and match facial matching to, you know, an existing database and say, okay, this person is now being attacked in this.
[19:09.930 --> 19:17.350]  You know, and I think the way that these things evolve, that's another change from the past.
[19:19.330 --> 19:29.330]  But if you go on social media right now and look at what people are saying about Trump getting shot in Butler, Pennsylvania, there's a lot of false flag accusations.
[19:29.330 --> 19:34.470]  There's a lot of, you know, funny memes as well, but this is a serious thing, right?
[19:36.030 --> 19:49.070]  No matter what you feel about the former president, and I'm not a fan of his, political violence at this level is something that, for one thing, it's illegal.
[19:49.710 --> 19:54.490]  The other thing is, it's, you know, a harbinger of potentially worse things to come.
[19:56.810 --> 20:03.210]  And, yes, it might be funny on one level, but it eventually leads to, like, actual conspiracy theories.
[20:03.510 --> 20:07.850]  Nobody took QAnon seriously until, what, 15% of the U.S.
[20:07.950 --> 20:10.490]  population actually believes the conspiracy theory, right?
[20:10.490 --> 20:11.630]  A concern.
[20:12.250 --> 20:16.870]  And then the final thing is, sometimes the call is coming from inside the house.
[20:18.170 --> 20:22.430]  We get a lot of questions and we don't work directly with law enforcement or anything like that.
[20:22.610 --> 20:28.350]  Sometimes if we see a direct threat, we will let people know because, like, we have a responsibility to do that.
[20:29.870 --> 20:42.210]  But for many of these companies, for many of the organizations we work with, it's sometimes their own employees who are leaking company information, who are going on to 4chan and, you know, self-radicalizing, right?
[20:42.310 --> 20:43.770]  It's coming from inside the house.
[20:43.970 --> 20:49.430]  And so it's not like you can just go and monitor this and leave it up to the government to fix.
[20:49.630 --> 20:55.970]  There's also, I think, a deeper social problem that's happening that needs to be better understood.
[20:59.050 --> 21:04.010]  You know, and these are some of the things that we try to educate our clients on, the actions that they take.
[21:04.690 --> 21:11.530]  I'll just point out that I know the title of this talk was talking a lot about A.I., you know, something that we've been working on.
[21:11.630 --> 21:12.930]  I came out of the human rights community.
[21:13.770 --> 21:25.210]  I was initially building a hate speech detector using Burt and Roberta bi-directional, you know, coding technologies to build a hate speech model.
[21:25.210 --> 21:32.270]  That led us to basically seeing the run-up to Jan 6, because we were monitoring Telegram.
[21:32.370 --> 21:37.350]  And we were warning basically anybody who would listen, like, the chatter is different this time.
[21:37.510 --> 21:38.150]  Something's happening.
[21:38.290 --> 21:39.050]  It's much more specific.
[21:39.950 --> 21:44.050]  There's videos of people going from Florida to D.C., right?
[21:44.130 --> 21:47.950]  It's just a... it's a different quality to the chatter.
[21:48.450 --> 21:49.910]  And no one would really listen to us.
[21:50.310 --> 21:57.610]  But we took that technology to combat what we saw as, like, one of the biggest threats that's happening, right?
[21:57.750 --> 22:01.830]  Because, yes, like, climate change may doom us all.
[22:03.150 --> 22:05.590]  You know, there's lots of other problems happening in the world.
[22:05.590 --> 22:20.650]  But if we can't agree on truth and we can't agree on, you know, what's okay to say on these platforms, I think that really causes us major problems for how we move forward as a society, right?
[22:20.810 --> 22:26.550]  Because it's really hard to have conversations with your neighbors if you can't even agree on basic facts.
[22:27.270 --> 22:35.330]  So with that, I'm going to reserve the rest of my time for questions and just thank you all for coming today and listening.
[22:36.130 --> 22:36.830]  Appreciate it.
[22:46.860 --> 22:49.460]  Oh, and if you have questions, you can just come to the front.
[22:50.960 --> 22:52.120]  If you don't, I can leave.
[22:56.940 --> 22:58.860]  Man, nothing I said was controversial at all.
[23:00.980 --> 23:02.380]  Everybody just agrees with everything?
[23:07.520 --> 23:09.740]  A lot of Reddit, a lot of 4chan.
[23:12.660 --> 23:16.240]  The 4chaners are basically saying, stop jerking off the porn.
[23:16.300 --> 23:17.880]  Our leader has just been shot.
[23:18.220 --> 23:23.200]  And then, you know, the Redditors are saying something along the lines of, I can't believe that guy missed.
[23:26.340 --> 23:26.880]  Yes.
[23:41.260 --> 23:43.560]  Yeah, we're not campaigning for censorship.
[23:44.200 --> 23:49.380]  That's, I think, a misunderstanding of what we do, first of all.
[23:51.240 --> 23:56.120]  Second, I would say there are First Amendment protected activities, certainly.
[23:56.580 --> 24:00.500]  But a direct death threat against somebody is not actually First Amendment protected.
[24:01.760 --> 24:08.960]  It's recognized by the Supreme Court via Brandenburg that, you know, you can't shout fire in a crowded theater, right?
[24:09.060 --> 24:09.680]  That's incitement.
[24:10.280 --> 24:25.640]  And so I would say that, like, a lot of the stuff that we see, you know, in terms of, like, I want to kill Supreme Court justice, or I want to kill Dr. Fauci, or I want to kill Oprah, that really doesn't have a place to live online.
[24:25.640 --> 24:26.220]  It shouldn't.
[24:27.380 --> 24:34.640]  I think we should agree that political violence is probably not the right solution.
[24:34.880 --> 24:38.680]  But you're free to say that you don't agree, right?
[24:39.260 --> 24:40.500]  That's First Amendment protected.
[24:46.110 --> 24:46.670]  Okay.
[25:07.490 --> 25:09.050]  I know it's really late.
[25:09.190 --> 25:10.030]  It's like 8 p.m., right?
[25:11.110 --> 25:11.350]  Check.
[25:11.830 --> 25:12.170]  All right.
[25:12.310 --> 25:12.710]  Excellent.
[25:12.710 --> 25:15.010]  So you're collecting a lot of important information.
[25:15.690 --> 25:16.410]  Two parts.
[25:16.610 --> 25:20.550]  Like, who are you notifying, like, when you see major threats?
[25:20.590 --> 25:23.390]  Like, is that the product that you're offering?
[25:24.730 --> 25:25.790]  Does that make sense?
[25:26.090 --> 25:26.390]  Yeah.
[25:26.810 --> 25:28.450]  So to be clear, we...
[25:29.410 --> 25:31.330]  Our company has software.
[25:32.150 --> 25:39.890]  Our clients access the content, and they're the ones doing, for example, we work with a number of different major anti-hater organizations in the U.S.
[25:39.890 --> 25:40.390]  and abroad.
[25:40.390 --> 25:43.090]  And they're the ones doing, like, law enforcement notifications.
[25:43.310 --> 25:44.210]  We're not the ones doing it.
[25:44.310 --> 25:47.870]  Now, we'll often be asked, hey, are you seeing this?
[25:47.970 --> 25:48.170]  Right?
[25:48.250 --> 25:50.410]  And then I'll go log on to the platform and take a look.
[25:51.350 --> 25:59.030]  And in cases where we do see a direct threat, like, we saw one against President Biden right before the inauguration.
[26:00.750 --> 26:05.950]  Basically, I asked friends if they knew somebody at the Secret Service, so I could get that over to them.
[26:07.030 --> 26:13.490]  And next, is there a potential that government will come and demand your data from you?
[26:14.630 --> 26:15.870]  That's a good question.
[26:16.570 --> 26:17.990]  I don't think so.
[26:18.210 --> 26:20.090]  We only collect open data.
[26:20.630 --> 26:24.610]  So anything that we have, the platforms themselves should have.
[26:25.430 --> 26:35.410]  We do sometimes get asked by, like, a community organization, for example, something got to beat it off of 4chan if we have it, because it's no longer accessible there.
[26:35.590 --> 26:40.590]  And we collect a lot of data from 4chan, so we retain that and are able to provide that.
[26:40.590 --> 26:41.870]  So, yeah.
[26:42.690 --> 26:43.050]  Yeah.
[26:43.190 --> 26:46.570]  And, you know, again, we don't work directly with law enforcement.
[26:46.850 --> 26:47.470]  Like, it's just...
[26:47.470 --> 26:54.750]  If you go look at our website or our terms of service, things like that, acceptable use policy, we're very, very strict about who we bring on as clients.
[26:55.530 --> 27:01.450]  And it's absolutely prohibited, and we reserve the right to audit, and it's part of our contract.
[27:01.690 --> 27:04.170]  Like, we can go and look at other people's projects, basically.
[27:04.910 --> 27:09.390]  You can't monitor civil rights organizations, protected human rights organizations.
[27:09.390 --> 27:10.830]  We just don't allow it, so.
[27:14.120 --> 27:14.640]  Yeah.
[27:14.800 --> 27:17.180]  My question mainly pertained to mental health.
[27:17.600 --> 27:27.060]  Having to dig through some of the most, like, repugnant and nefarious corners of the Internet obviously takes extreme toll on your mental health.
[27:27.240 --> 27:31.480]  Like, is that something that you need to, you know, doubly prioritize?
[27:31.480 --> 27:40.560]  Is there anything, you know, routines or anything you can recommend for those of us, you know, that have to monitor some of that as part of our jobs or research?
[27:41.480 --> 27:53.340]  Yeah, I mean, our last company retreat, we did two hours with specialists that work with the response teams at Meta and places like that where they're seeing really terrible stuff as well.
[27:55.220 --> 27:58.160]  I don't know if I got a ton out of that two hours, to be honest.
[27:58.660 --> 28:00.900]  It's also, like, the first time I'm ever taking medication.
[28:01.600 --> 28:11.020]  So, like, you know, and I mean, you look at my bio, I did two combat tours in Iraq, so, you know, it's not, I'm not a stranger to stress, but it's a different level.
[28:12.460 --> 28:19.100]  I think, you know, do everything to keep yourself healthy, get regular exercise, eat right, like, get sleep.
[28:19.360 --> 28:24.900]  You got to do all those things at the baseline and then recognize your own limits and what you're comfortable with.
[28:24.980 --> 28:37.080]  Like, it's one of the things that we talked to everybody that we hired as a company was, like, me spending a good 15 minutes talking to them and trying to convince them out of joining the company and saying, like, you're going to be encountering terrible,
[28:37.240 --> 28:38.940]  terrible stuff all the time.
[28:39.540 --> 28:49.180]  And, I mean, one of the things that we did in the company is built a special fine-tuned model to find things like KKK hoods, for example.
[28:50.000 --> 28:55.620]  And I spent a lot of time drawing bounding boxes around KKK hoods to fine-tune the model.
[28:55.900 --> 28:59.100]  And I didn't think that that would affect me deeply.
[29:00.040 --> 29:08.800]  But when you, you know, look through several hundred photos of lynching, like, you know, it's just not, it's not great.
[29:09.700 --> 29:12.160]  So, and look, that's why we have AI, right?
[29:12.400 --> 29:14.560]  The AI can help you sift through that.
[29:15.380 --> 29:18.960]  Media in our platform, you have the option of blurring all of it, right?
[29:19.100 --> 29:21.520]  So if you don't want to see any of it, you can just blur all of it.
[29:22.160 --> 29:24.320]  And the less that you counter it, I think the better.
[29:24.860 --> 29:27.100]  It's just, yes, it exists.
[29:28.280 --> 29:31.320]  And it's the hidden part of the Internet that nobody acknowledges.
[29:31.560 --> 29:41.680]  Why you don't regularly see child pornography on Twitter or on Facebook, it's because there are humans reviewing the AI detections and removing that from those platforms, right?
[29:41.780 --> 29:44.160]  That stuff all exists, it's just being hidden from you.
[29:44.380 --> 29:46.460]  And there's human costs to it, absolutely.
[29:47.400 --> 29:50.260]  For your work, we appreciate you and thank you for the talk.
[29:54.050 --> 30:04.250]  It would be naive for you to believe that your stuff won't be eventually subpoenaed by law enforcement and by private actors.
[30:06.690 --> 30:11.250]  I work at the Internet Archive and we get requests all the time.
[30:12.390 --> 30:23.950]  And our people often have to go to court, not universally, and vouch for the accuracy or the integrity of documents that we preserve.
[30:26.410 --> 30:27.310]  Thanks for that.
[30:36.850 --> 30:48.450]  Do you believe that there's a chance that by restricting content like this to these major platforms, it might cause people people to go to even more drastic lengths to get their hateful messages out there?
[30:49.670 --> 30:51.030]  I'm sorry, could you repeat that question?
[30:51.530 --> 31:06.350]  Do you think that it is possible that by restricting access to these more open platforms, people might go to more drastic lengths, which would in turn cause more damage in order to try and get their hateful messages out there?
[31:07.610 --> 31:09.070]  I mean, I think that's already happening.
[31:09.370 --> 31:15.810]  I think people who are, you know, trying to kidnap the governor of Michigan, for example, right, they were using Signal.
[31:19.570 --> 31:23.350]  I don't... I don't see a way to, you know, stop people from using Signal Chat.
[31:23.910 --> 31:32.410]  But I think the question really is, in order for... and look, there's a ton of nefarious activity happening at Dark Web as well, right?
[31:33.230 --> 31:40.130]  I think the question is, what happens when you have a bigger microphone to broadcast your intentions and get people to join in?
[31:40.210 --> 31:42.090]  That's how you get events like January 6th.
[31:42.150 --> 31:43.830]  That's how you get events like what happened in Ottawa.
[31:49.140 --> 31:57.900]  These sites that are popping up now, a lot of them came up because the folks got basically kicked off of Twitter and those places, right?
[31:58.120 --> 32:04.160]  Like, True Social exists because of, you know, some of the enforcement actions taken by Twitter.
[32:04.180 --> 32:07.740]  But I think one of the things to recognize is that True Social is a much smaller platform.
[32:07.960 --> 32:14.160]  You just get less engagement and less ability to broadcast, you know, terrible stuff.
[32:16.280 --> 32:16.960]  Yeah.
[32:21.850 --> 32:28.890]  So, I seem to understand that the main goal here is public safety.
[32:29.230 --> 32:29.650]  Yes.
[32:29.650 --> 32:54.010]  So, I was wondering, in a sense, as a private actor that develops models that then run continuously, you take responsibilities, both like that, in a sense, going in a mix from the technical to the definition of a task and to what kind of cases are covered.
[32:54.010 --> 33:12.330]  So, my question is, in a sense, what kind of responsibilities do you take, well, technically, politically, and legally, when you give your models, your software to other institutions who use it?
[33:12.330 --> 33:12.930]  Yeah.
[33:13.790 --> 33:16.890]  I mean, I think the first thing to recognize is that the AI is not perfect.
[33:17.390 --> 33:20.190]  And I don't think anybody expects the AI to be perfect.
[33:21.830 --> 33:32.390]  Right now, if you go and look at extremely violent posts related to the Trump shooting, you can find a bunch in our platform right now.
[33:33.210 --> 33:34.970]  Some of them are going to be sarcastic.
[33:35.150 --> 33:39.250]  Some of them are going to be basically false detections.
[33:41.670 --> 33:48.730]  But you've necked down the universe of things you have to look through, rather than just doing a strict bullying search.
[33:50.550 --> 33:54.610]  And you're able to sort by what the AI thinks is extremely violent.
[33:56.850 --> 34:00.550]  Yeah, I think our clients are very realistic about how the AI performs.
[34:01.310 --> 34:03.310]  And we are very realistic about that, too.
[34:03.510 --> 34:05.370]  And I think it's something that I say all the time.
[34:05.610 --> 34:09.690]  If somebody comes around and says, like, oh, the AI can do everything, like, that's just not true.
[34:09.690 --> 34:27.330]  The best models out there running, you know, the latest and greatest transformer technology, with, you know, immense amount of fine tuning happening, even that's only going to be, I don't know, maybe 85% accurate.
[34:27.930 --> 34:30.730]  Like, that's, that's a great model, in my mind.
[34:31.230 --> 34:37.370]  You know, a first cut, getting something 68%, 71% accurate, like, you're pretty happy.
[34:39.350 --> 34:46.190]  But, when you're talking about millions and millions of posts a day, I just don't see a way of looking through all of that without the assistance of AI.
[34:46.370 --> 34:48.830]  You have to have something that helps them through it.
[34:50.510 --> 34:51.510]  It's not going to be perfect.
[34:52.870 --> 34:53.470]  Well, thanks.