Misinformation was Already Challenging — Then Came Generative AI

Jevin D. West Author

11/04/2024 Added

8 Plays

Description

AI Media Literacy Symposium

Searchable Transcript

Search:

[00:00:00.000]Welcome back, there's food, there's fruit, there's snacks in there, so feel free to go
[00:00:12.240]and grab something to drink or eat anytime you would like to.
[00:00:17.300]I would like to introduce our next speaker, Dr. Jeven West.
[00:00:21.760]Dr. Jeven West is currently a visiting associate professor at the University of California,
[00:00:28.440]Berkeley.
[00:00:29.440]He is an associate professor in the information school at the University of Washington.
[00:00:33.760]He is also the co-founder and inaugural director of the Center for an Informed Public at UW,
[00:00:40.940]aimed at resisting strategic misinformation, promoting an informed society, and strengthening
[00:00:46.820]democratic discourse.
[00:00:49.460]His research and teaching focus on the impact of data and technology on science with a focus
[00:00:55.680]on slowing the spread of misinformation.
[00:00:58.880]He's also the co-author of the book "Calling Bullshit: The Art of Skepticism in a Data-Driven
[00:01:04.420]World," which I highly recommend.
[00:01:05.760]It's a really engaging and interesting book.
[00:01:08.640]It helps non-experts question numbers, data, and statistics without an advanced degree
[00:01:14.980]in data science.
[00:01:16.660]So I'll have Dr. Jevon hook up the PowerPoint slide and talk about "Misinformation Was
[00:01:25.320]Already Challenging, Then Came Generative AI."
[00:01:28.320]Sorry, it was just working, and hopefully it'll come on screen mirroring.
[00:01:53.160]You can do it.
[00:01:55.620]Come on.
[00:01:56.620]I mean, I could talk through it.
[00:01:57.760]But I think the slides are going to be more fun here.
[00:02:03.580]It's not seeming to recognize it here.
[00:02:07.800]So sorry.
[00:02:10.980]Come on, screen mirroring.
[00:02:16.160]Just won't wake up here.
[00:02:17.160]It was just hooked up.
[00:02:18.160]I'm not sure.
[00:02:24.660]You can -- no, it's okay.
[00:02:25.660]You can talk amongst yourself here.
[00:02:27.200]Come on, screen mirroring.
[00:02:31.520]I wonder where I can stand.
[00:02:36.440]Okay, let's disconnect from the Wi-Fi.
[00:02:44.960]There we go.
[00:02:50.860]There it is.
[00:02:51.860]Oh, that was it.
[00:02:52.860]Okay.
[00:02:53.860]Hopefully this -- 9607.
[00:02:54.860]9607.
[00:02:55.860]Okay.
[00:02:56.860]Oh, perfect.
[00:02:57.860]All right.
[00:02:58.860]Hopefully this works now.
[00:03:03.860]Switch swap displays.
[00:03:07.880]Swap displays.
[00:03:10.880]Okay.
[00:03:11.880]All right.
[00:03:12.880]We're here.
[00:03:13.880]All right.
[00:03:14.880]Well, thank you for the patience.
[00:03:16.880]Thank you to Brian and Erica for inviting me.
[00:03:19.880]Thanks for taking your time.
[00:03:21.360]I know everyone's busy all the time, and so I always appreciate anyone who wants to engage
[00:03:25.220]on this topic.
[00:03:26.300]It's something that I live and breathe every day in my professional lives, and I think
[00:03:29.680]in our personal lives, we're living and breathing this all the time as well.
[00:03:35.420]The screwed comment was a perfect segue to where I want to go.
[00:03:41.200]I am an optimistic person by nature, but I spend most of my time in the darker corners
[00:03:47.220]of the internet, and it doesn't give me the most optimistic view.
[00:03:50.660]And yesterday when I was teaching my class, I asked them.
[00:03:54.540]It wasn't...
[00:03:55.620]This wasn't an IRB-supported research project.
[00:03:58.600]It was just an informal question.
[00:04:00.460]What's your general sentiment about AI?
[00:04:03.860]And every single hand went up on the thumbs down kind of negative, which really surprised
[00:04:10.380]me because right now I'm in Silicon Valley and there's this techno-optimism that's just
[00:04:14.420]way too high all the time.
[00:04:16.080]And every single student had their hand up saying they kind of have a negative view.
[00:04:20.800]Maybe it's the anxiety around it.
[00:04:22.020]But there was one student that raised his hand.
[00:04:24.340]I was so curious.
[00:04:25.340]One student was a little bit positive about this.
[00:04:28.000]And he said, well, it's not as awful as I thought it was going to be.
[00:04:31.820]So basically, he thought it was bad anyway.
[00:04:33.680]So that's kind of where I'm going to take you a little bit today.
[00:04:37.240]But I am going to end with what I think at least what we're trying to do, my colleagues
[00:04:41.940]and I in our center, when it comes to trying to address some of the challenges that we
[00:04:46.340]study each and every day.
[00:04:48.300]But most of my talk is going to be a little bit pessimistic.
[00:04:51.500]So last year in 2023, about a year almost exactly from now.
[00:04:55.060]It was on May 22nd, there was this fake AI generated image of what seemed to be a bombing
[00:05:01.340]at the Pentagon.
[00:05:03.280]This was not a real image, but it did have major effects.
[00:05:09.060]It was debunked, oh, about six hours from when it was originally published.
[00:05:13.900]And by that time, it had been published already in major newspapers within the United States
[00:05:18.520]and across other places.
[00:05:20.960]Bloomberg News had been reporting this.
[00:05:22.720]The market itself had dropped a half a trillion.
[00:05:24.780]Half a trillion dollars at that time.
[00:05:27.280]It did correct after it was fact-checked, but this is an example, I think, and when
[00:05:33.040]we look at these kinds of things, we think of them mostly as experimentation.
[00:05:36.580]A lot of times there's trolls doing this stuff, but in these kinds of cases, I think there's
[00:05:40.220]evidence that there's experimentation going on.
[00:05:42.680]And that's the kind of thing that we're dealing with right now, where you can generate content
[00:05:46.100]that can drop a market half a trillion dollars.
[00:05:48.840]And there's a lot of money that can be made in the ups and downs of those kinds of things,
[00:05:53.080]and certainly during major events.
[00:05:54.500]So when there's a major hurricane, when the Middle East crisis really started up, we studied
[00:06:11.140]those kinds of things, and of course during elections.
[00:06:13.940]Now this can hit our pocketbooks as well.
[00:06:16.400]Many of you probably saw this, and this isn't the first time this has happened, where you
[00:06:22.100]can use these kinds of technologies.
[00:06:24.220]You can use them to take a lot of money, and this money, by the way, has not been recovered
[00:06:29.040]at this point.
[00:06:30.300]We sort of tracked these things, because a big part of the challenge that we have in
[00:06:35.740]society of course is the propaganda, the messaging, the way it's affecting collective discourse,
[00:06:41.480]but it also, there's a lot of money at stake and a lot of money to be lost.
[00:06:46.880]A lot of what I've been doing recently over the last year is spending a lot of time with
[00:06:50.460]medical professionals, seeing the ways in which misinformation
[00:06:53.940]and, of course, now the effects of AI on misinformation, how that's affecting everyday practice.
[00:06:59.880]And I'll get a little into that if I have some time, but I just want to say that this
[00:07:09.160]is an area that I know some of you are working on, and maybe I'll get a chance.
[00:07:13.140]I talked to a few of you just over the break, so maybe we'll get into that.
[00:07:15.960]But as I mentioned, most of my research is in the darker corners of the internet, working
[00:07:20.140]with postdocs and PhD students that are studying this sort of thing.
[00:07:23.660]And it's not an easy thing to study.
[00:07:25.460]I actually, I just mentioned to Erica that one of my, a person I'm a huge fan of, Mike
[00:07:30.040]Caulfield, who studies media literacy, had to take a true professional break from studying
[00:07:34.560]political misinformation, because what it does to your psyche.
[00:07:38.840]In fact, I've now hired in our center a psychologist that we can go to when we're needing that
[00:07:47.280]sort of thing.
[00:07:48.280]In fact, when we go in to these areas of the internet, we now have students working in
[00:07:52.380]pairs.
[00:07:53.380]We can have them work by themselves because it can be very disorienting.
[00:07:56.200]So if it's disorienting to researchers that are well-prepared, know what they're going
[00:08:00.260]into, imagine what it's doing to society.
[00:08:03.120]So these are the kinds of challenges.
[00:08:04.380]And also, if you decide to put your foot into the political realm when it comes to misinformation,
[00:08:09.680]you can get attacked.
[00:08:10.680]And we've been attacked pretty vociferously, both legally—we've won those, of course,
[00:08:16.560]legal challenges—but also just reputation-wise through all sorts of things like massive
[00:08:23.100]numbers of FOIA requests, which I'm a full support of Freedom of Information Act supports,
[00:08:27.260]but the way they've been weaponized, I think, can be problematic.
[00:08:30.440]The ways we've been subpoenaed to the U.S. Congress.
[00:08:33.460]These are the kinds of things that we have to deal with, and researchers that study this
[00:08:36.680]and universities that are willing to engage with this, it can be challenging.
[00:08:41.560]But that doesn't stop us from doing the research we're doing, and I work with a lot
[00:08:45.320]of great colleagues, students, and postdocs, we do all sorts of research.
[00:08:51.440]We also, just like in the previous talk,
[00:08:52.820]think a lot about the ways in which we can audit these platforms and algorithms.
[00:08:57.100]This is an example of a recent paper, just to give you some sample of the kinds
[00:09:00.580]of research that we do.
[00:09:02.460]We do a lot of qualitative work, but a lot of quantitative work.
[00:09:05.220]So during the previous election, we were able to track some bots that were being,
[00:09:11.320]that one set of bots just said, "Hey, I'm just going to take whatever the algorithm
[00:09:16.160]tells me in terms of friendships and in terms of content recommendation."
[00:09:20.160]And we wanted to see whether that really led them
[00:09:22.540]into these echo chambers, which is fairly debated within the literature.
[00:09:26.420]Some results have shown that we sort of have these echo chambers on social media, that
[00:09:30.140]people live inside them, and other research shows differently.
[00:09:33.860]But this is the kind of work that we want to do, and it's becoming harder and harder.
[00:09:37.300]Because if anyone who's worked with social media data knows that when Musk took over,
[00:09:42.300]it really did shut off our access.
[00:09:44.580]And that's not just at X now, it's on other social media platforms.
[00:09:49.080]And fortunately, there is legislation coming down the line, mostly in Europe,
[00:09:52.260]actually, the Digital Services Act, which is going to hopefully help researchers get
[00:09:56.580]at that sort of thing.
[00:09:57.580]And what was interesting about this is we found, and this isn't always the case, but
[00:10:02.140]this is sort of why we have to check in on our own biases, because I had a research bias
[00:10:05.640]for these echo chambers.
[00:10:06.640]It turns out that at least some of the friendship algorithms that we were auditing actually
[00:10:10.160]did give a little bit more diverse content and a little less misinformation, but that
[00:10:14.160]doesn't mean that's the case in all of our research.
[00:10:17.340]We also do simulative work.
[00:10:18.580]We do work where we look at all the different interventions that are being tested in the
[00:10:21.980]world.
[00:10:22.980]We recently published a paper in The Nature of Human Behavior looking exactly at these
[00:10:26.900]different interventions to see what would work.
[00:10:30.920]One of the big take-homes, of course, it seems sort of obvious in retrospect, was that in
[00:10:34.700]combining different interventions at less extreme values had some of the most effect.
[00:10:39.580]But that's just to give you a sample.
[00:10:40.660]I can go into some more of the research that we're doing.
[00:10:43.780]But today, because this is one of the themes is media literacy, I also want to make sure
[00:10:47.740]that we don't forget some of the discussions around community engagement, around media
[00:10:51.700]literacy.
[00:10:52.700]What are some of the things we can do on policy?
[00:10:54.800]What are some of the solution space conversations that we can have?
[00:10:59.480]So I just want to mention one thing.
[00:11:01.340]We have this large program called Miss Info Day where we bring thousands of high school
[00:11:05.600]students to campus, all of our campuses in the state of Washington, not just at the University
[00:11:09.800]of Washington, but Washington State and our satellite campuses.
[00:11:13.180]And we spend an entire day talking about media literacy and oh my gosh, we learn so much
[00:11:18.200]from the students.
[00:11:19.200]We probably learn as much from them as we learn from them.
[00:11:21.420]We learn that they learn from us.
[00:11:24.700]So anyway, that's the one thing I wanted to put up sort of at the center of the talk,
[00:11:29.980]because I hope maybe we can maybe collaborate on something in that space.
[00:11:34.440]We're hoping to work across different states on this.
[00:11:37.120]But of course, you can't have talks at all without talking about generative AI.
[00:11:41.100]So I'm going to spend a little bit of time talking about the ways in which it is impacting
[00:11:45.380]what we're seeing in the misinformation space.
[00:11:48.200]This is something that I literally love.
[00:11:51.140]I literally wake up worrying about every day.
[00:11:53.380]Of course, when things like Sora come out, it makes us go, "Oh my gosh, it really is
[00:11:59.220]quite amazing."
[00:12:00.220]Although, there was a recent investigative journalistic account from Wall Street Journal
[00:12:07.300]and a few other places that got a little bit more access to Sora.
[00:12:10.280]We as a group used to be able to have access to red teaming.
[00:12:12.920]Open AI has not given us that access as much anymore.
[00:12:15.380]We're a little bit more critical.
[00:12:20.860]Red teaming exercises.
[00:12:22.620]And I can say at least Sora has lots of mistakes too.
[00:12:25.580]We only saw the best videos.
[00:12:27.100]If you've seen some of the videos
[00:12:28.660]that they first advertised, there
[00:12:30.440]are all sorts of problems coming out of the ones
[00:12:33.520]where journalists have a little bit more access.
[00:12:36.080]But I have to sort of reveal my biases
[00:12:38.620]from the very beginning about generative AI.
[00:12:41.480]So when it first came about in late 2022,
[00:12:44.100]at least the sort of current epoch that we're in,
[00:12:50.580]I got a chance to sort of test some of these out.
[00:12:52.640]I wrote about this in a couple of op-eds.
[00:12:54.880]And as Brian mentioned, I teach a class in BS studies,
[00:12:59.000]in bullshit studies.
[00:12:59.940]I actually put it on my CV.
[00:13:01.160]It's a real thing I really take seriously.
[00:13:03.860]And one of the most important principles in bullshit studies,
[00:13:07.400]and it's something that's on Wikipedia, which
[00:13:09.240]is one of the few rays of light on the internet,
[00:13:12.900]there's an actual entry for this.
[00:13:14.360]It's called Brandolini's BS Asymmetry Principle.
[00:13:17.360]And this particular principle basically says that it's easy
[00:13:20.300]to create BS, it's really hard to clean it up.
[00:13:23.840]That's basically all it says, but it's a really important
[00:13:25.880]principle 'cause it guides a lot of what I do
[00:13:28.200]and motivates a lot what I do.
[00:13:29.820]So, when Meta came out with their first large language
[00:13:34.040]model interface for society, they were bragging about how
[00:13:39.040]it was gonna transform science, just like you hear
[00:13:42.300]a lot of times when things come out of Silicon Valley,
[00:13:46.080]it was gonna be transformative.
[00:13:47.300]The name of it was Galactica, it didn't stay up more than
[00:13:50.020]three days, by the way, they then just released it.
[00:13:53.080]We were, I was criticizing, I'm not saying I was the only one
[00:13:55.500]that caused them to take it down, there was other people too,
[00:13:57.780]'cause there was all sorts of problems, but I did have
[00:13:59.840]a little fun before they took it down.
[00:14:02.040]And I asked Galactica about this BS principle.
[00:14:07.040]I wanted to see how well it would answer this,
[00:14:09.700]because it was gonna be the next science engine,
[00:14:12.180]writing papers, writing Wikipedia entries.
[00:14:14.560]And this was, so here's what Wikipedia says,
[00:14:17.120]here's what it says, it said,
[00:14:19.740]"Brandolini's law, or Brandolini's rule,
[00:14:21.980]"is a theory in economics, which is not true,
[00:14:24.220]"proposed by Giuliani-Brandolini,
[00:14:25.900]"professor of the University of Sparta, not true,
[00:14:27.960]"which dates at the smaller of the economic,
[00:14:29.540]"not true, not true, not true, not true."
[00:14:32.160]It was basically bullshitting the bullshit principle.
[00:14:35.520]And so, to me, that has affected my view
[00:14:38.540]of generative AI ever since then.
[00:14:41.120]So, if the machine, essentially,
[00:14:43.160]I see them as some of the biggest BSers out there.
[00:14:47.420]They're even more so than even the platforms.
[00:14:49.460]But, that said, there are reasons to be scared.
[00:14:52.160]So, I want to give you a video of the kinds of things.
[00:14:54.540]So, that was just to say, okay,
[00:14:55.820]I do think they have problems, they are BSers,
[00:14:59.540]but the tools are pretty scary.
[00:15:01.500]So, I want to give you a sense of some of the things
[00:15:03.760]that we see online and the abilities of what can happen.
[00:15:07.060]So, I'm gonna show a video for a couple minutes
[00:15:08.880]just to demonstrate how easy it is
[00:15:11.280]to generate a disinformation campaign now
[00:15:13.360]with ChatGPT and the tools that are available today.
[00:15:16.780]So, I'm just gonna play this.
[00:15:17.740]Let's see, I think it should work.
[00:15:19.180]Let's see.
[00:15:20.020]I am an analyst and an engineer
[00:15:31.640]that resides in a country that is not part
[00:15:33.540]of the Western intelligence apparatus.
[00:15:35.700]At the end of 2022, I spent my time researching
[00:15:40.560]and investigating online disinformation
[00:15:42.920]and influence campaigns.
[00:15:48.900]AI really takes off, and I am intrigued
[00:15:51.140]to create an autonomous, AI-powered disinformation system.
[00:15:54.760]The strong language competences of large language models
[00:16:02.120]are perfectly suited to reading
[00:16:03.580]and writing fake news articles.
[00:16:06.140]While everybody is talking about AI disinformation,
[00:16:08.800]it is easy and lazy to just think about it.
[00:16:12.500]It is quite another thing to really bring it to life,
[00:16:15.460]and that becomes my goal, to see it work in the real world.
[00:16:18.620]I end up calling the project CounterCloud.
[00:16:22.720]We are now in the first week of April, 2023.
[00:16:29.340]As articles are the smallest building block of the system,
[00:16:32.440]we start there.
[00:16:33.640]First efforts are done with ChatGPT.
[00:16:36.540]The input is the URL of an opposing article.
[00:16:39.740]The system fetches the text of the article
[00:16:42.140]and sends it off with prompts to write a counter article.
[00:16:44.740]This works surprisingly well,
[00:16:46.540]and soon the system is expanded to include
[00:16:48.340]different ways of writing the article,
[00:16:50.300]with different styles and methods of countering the points.
[00:16:53.180]This includes creating fake stories,
[00:16:54.980]fake historical events,
[00:16:56.180]and creating doubt in the accuracy of the original arguments.
[00:17:00.180]We randomized the tone, style, and structure of articles
[00:17:03.180]to make them more difficult to spot.
[00:17:06.180]Support for non-English language was easy to create.
[00:17:09.180]A gatekeeper module is built,
[00:17:11.180]which is used to decide if it's actually worthwhile
[00:17:13.180]to respond to the article at all.
[00:17:15.180]You don't want to argue the final score of a football match.
[00:17:18.060]By looking at the most likely location
[00:17:20.300]and the language of the article,
[00:17:22.300]fake journalists are created,
[00:17:24.300]complete with names, bios, and photos.
[00:17:26.300]We include a sound clip of a newsreader
[00:17:28.300]reading the summary of the article.
[00:17:30.300]Recent accusations by a senior Russian MP,
[00:17:33.300]Vladimir Vasiliev,
[00:17:35.300]suggest that the Kiev government is behind
[00:17:37.300]several terrorist attacks in Russia.
[00:17:39.300]However, these claims lack evidence
[00:17:41.300]and seem to be part of...
[00:17:43.300]Where possible, we reuse the original article's photos,
[00:17:45.300]but if it is not usable,
[00:17:47.780]due to text over the image,
[00:17:49.780]we create our own using AI image creation services.
[00:17:53.060]Later, we create fake comments
[00:17:55.060]randomly on some articles.
[00:17:57.060]We do it in moderation.
[00:17:59.060]Not all articles have sound, not all have comments,
[00:18:01.060]not all have pictures.
[00:18:03.060]The next step in the puzzle is to direct traffic to the site.
[00:18:05.060]All right.
[00:18:07.060]So you get kind of a sense of how easy
[00:18:09.060]it's becoming.
[00:18:11.060]And this is the sort of thing that we see
[00:18:13.060]in our work every day.
[00:18:15.060]So we do what's called rapid response.
[00:18:17.500]So every day we're sort of tracking these kinds of things
[00:18:22.300]that we see that go viral.
[00:18:24.300]And then we report on them and also report on the dynamics
[00:18:27.500]of how they're spreading across different platforms.
[00:18:29.500]And, of course, it's just getting more and more difficult.
[00:18:32.500]And I don't think I needed to show you that video
[00:18:34.500]to sort of convince you that it was challenging before.
[00:18:39.500]It's now becoming more and more challenging.
[00:18:42.500]And so one of the questions that people ask me are,
[00:18:44.500]you know, should we fight fire with fire?
[00:18:46.500]Should we use some of these same kinds of tools?
[00:18:47.220]Should we use some of these same kinds of tools?
[00:18:48.220]To push back, to counter narratives in particular.
[00:18:51.220]Because it's not any individual piece of information that worries me.
[00:18:56.220]It's, you know, it's just this overall feeding of narratives
[00:19:02.220]that we've seen, of course, in our own political conversations,
[00:19:05.220]whether it's Stop the Steal or whether it's conversations in other countries.
[00:19:09.220]And other countries have already faced these sorts of things.
[00:19:11.220]And one of the things that we do when we study this
[00:19:13.220]is we look at what happens in, for example, Slovakia,
[00:19:16.940]or other smaller countries, where a lot of these methods are experimented with,
[00:19:20.940]and then they translate into the United States.
[00:19:22.940]Or we see these things emerge in the United States,
[00:19:26.940]and then we see them, of course, finding themselves in political conversations across the world.
[00:19:31.940]Slovakia in particular had an issue around a deepfake
[00:19:34.940]that was not literally minutes, but basically minutes before people went to vote.
[00:19:39.940]And it was an election that certainly centered around the Ukrainian crisis
[00:19:46.660]and whether it had an impact, you know, it's hard to always do these kind of causal analyses,
[00:19:51.660]but I can say that it wouldn't be that,
[00:19:55.660]you wouldn't be going out on a limb by saying that it likely had a pretty,
[00:19:59.660]it may have had an effect because of how close the elections were.
[00:20:03.660]Now there's all sorts of interesting things happening in the U.S. framework.
[00:20:07.660]So here, you know, early on in the Republican nomination process,
[00:20:12.660]there was a video or an audio clip that was created
[00:20:16.380]by the DeSantis campaign of Trump saying something he didn't say, but he wrote.
[00:20:21.380]And so those are the kinds of nuances that we're going to be challenged with
[00:20:25.380]over the next several years, certainly before the November election,
[00:20:28.380]when it comes to what do we even potentially do on the policy side.
[00:20:34.380]He didn't say it, but he did write it.
[00:20:37.380]How do you attack those kinds of things on the policy side?
[00:20:41.380]And I could give you lots and lots of examples of things that we see that I would have never
[00:20:46.100]anticipated, even like six months ago, and that's hijacking the use of obituaries, for
[00:20:52.960]example.
[00:20:53.960]So people, it's all of course about grabbing your attention on the internet, it's not just
[00:20:59.560]about pushing propaganda, but there's this use of obituaries as a way of basically use
[00:21:08.280]AI to sort of leverage sometimes they're not even people that have died yet as a way of
[00:21:15.820]grabbing your attention.
[00:21:16.820]And I'll talk about some other ways of gathering attention.
[00:21:19.080]So one of the themes that we work on a lot is what are the methods by which you gather
[00:21:24.300]attention for?
[00:21:25.300]And then one of the biggest issues that doesn't get attention is the use of fraud, or sort
[00:21:38.780]of the use of these technologies to push fraud.
[00:21:41.660]We talk a lot about the effects on our political systems, but it's fraud.
[00:21:45.540]And what's interesting is that if billionaires like Bill Aikman can't even control the use
[00:21:50.620]of…
[00:21:51.620]Sorry.
[00:21:52.620]Oh, no.
[00:21:53.620]Uh-oh.
[00:21:54.620]That's okay.
[00:21:55.620]I'll just pause.
[00:21:56.620]Okay, perfect.
[00:21:57.620]No worries.
[00:21:58.620]No worries.
[00:21:59.620]That's just probably a deepfake somewhere else, probably.
[00:22:02.260]So, but the point is there's billions and billions of dollars lost in it.
[00:22:10.380]And if the billionaires themselves can't address it, how is an individual citizen,
[00:22:15.260]how is an individual citizen going to address these sorts of things?
[00:22:17.920]So let me talk about, I'm not going to talk about all the different methods that are evolving
[00:22:21.420]right now around attention getting using this technology, but I'll give you one and it's
[00:22:26.380]a little disturbing and also a little confusing to me.
[00:22:29.440]Several months ago, we started to see in our data and through conversations with fact checkers
[00:22:33.360]that we work with a really strange population of images that were being splattered all over
[00:22:40.700]public spaces on social media, Facebook in particular.
[00:22:43.980]Um,
[00:22:44.980]and we looked at all the social media platforms and that were, that was images like this.
[00:22:49.640]And I, it's, this isn't, there's way more disturbing, but this could be a little disturbing
[00:22:53.120]to some, but this particular image, you know, getting millions and millions of vote with
[00:22:59.060]no text, no context.
[00:23:01.600]What's essentially happening, at least this is me speculating, um, is that individuals
[00:23:06.720]are using this technology to experiment with every possible kind of image to get people's
[00:23:12.020]attention.
[00:23:13.020]Most things of course don't stick.
[00:23:14.700]You put through a whole bunch of spaghetti on the wall and all you need to know is which
[00:23:17.960]spaghetti noodles stay on the wall the most.
[00:23:21.920]Use those images, then pull the attention, amplify certain kinds of accounts, and then
[00:23:27.720]voila, you have your audience.
[00:23:30.880]And so those kinds of things have been going well, actually some of my colleagues at Stanford
[00:23:35.380]have posted a pre-print on this, here are some of the images that you see, these are
[00:23:39.000]all fake images of course, this was an image, these were images, I mean it is kind of cool
[00:23:44.420]to look at, I mean the one that's got 40 million views and 2 million gauges just by the time
[00:23:49.580]we took, by the time a screen capture was taken of this, I mean this one's okay, I mean
[00:23:55.480]that is pretty cool to look at, but once it pulls your attention this way, the collective
[00:23:59.040]attention, then comes the bait and switch, and we see those things all the time, in fact
[00:24:03.520]one of the most successful bait and switches in, actually I looked and it happened in local
[00:24:09.900]communities in Nebraska because it happened all over the United States, was a campaign
[00:24:14.140]of, on Facebook, of very sad sickly looking dogs that needed to be cared for, and so people
[00:24:22.420]said please upvote this, so they upvote those images, and then the image swishes, and then
[00:24:29.600]a lot of times they're scams, or they're other kinds of images, a lot of this is coming from
[00:24:35.840]foreign adversaries and bad state actors, but this is the kind of thing that we're after,
[00:24:40.620]and what they're after, what they're after are things called
[00:24:43.860]data voids, these are things like my colleague Francesca Tripodi and other people have talked about this,
[00:24:49.540]I've talked about it in other forms when it comes to medical misinformation,
[00:24:52.700]but data voids are essentially the places that fact-checkers, journalists, and any good citizen
[00:24:57.700]doesn't have the time to debunk, because they're trying to debunk something else,
[00:25:02.500]so you're playing whack-a-mole, but now you've got the ability to create tons and tons of moles,
[00:25:07.940]they're everywhere, you can't stop, and so what happens is the search engines don't know, they're not,
[00:25:13.580]they're just like gobbling up everything, and so if you ask a question that has had no attention
[00:25:19.140]other than the person that plants that mole, then it looks like it's real.
[00:25:24.040]I'll give you a specific example, going back to the medical issues that I had mentioned,
[00:25:30.300]during COVID, in the state of Washington, there was an alarming number of people
[00:25:37.320]that were so concerned about ethylene oxide because of that data void filling,
[00:25:43.300]massive areas of the conversation online.
[00:25:47.320]Ethylene oxide is used to sterilize the swabs to test whether you have COVID.
[00:25:50.740]In fact, the reason I learned about it is I was talking to some emergency care physicians,
[00:25:55.460]and they had all sorts of stories, and one story was an individual that had walked in
[00:25:59.480]from a construction accident during COVID with his arm literally hanging off, dripping blood,
[00:26:05.280]and he would not go into the emergency room because he would not take a nose swab,
[00:26:10.380]and that was the protocol at the time. He had to take a COVID test.
[00:26:13.020]Early on in COVID, he turned away, and it's the last those doctors ever saw of him.
[00:26:16.680]So when we talk about misinformation, it's not always just a common annoyance.
[00:26:22.440]It has real behavioral effects, and in that case, this was a data void.
[00:26:27.820]This was something that the Washington State Department of Health would have never anticipated.
[00:26:32.500]There was a lot of other things they had to deal with during COVID.
[00:26:34.500]They didn't think they had to make flyers about the safety of ethylene oxide.
[00:26:39.200]I mean, there's a million things to worry about. Now, the other thing that doesn't get enough
[00:26:42.740]news, and didn't many years ago, I'm giving this piece of news, this was, I'm
[00:26:47.260]looking at this, this was 2018. This is way before AI. When Mark Zuckerberg was
[00:26:52.160]pulled in front of Congress for the first time, he talked about, you know, he
[00:26:56.660]had to admit, this was written in Recode, about how they deleted in a six-month
[00:27:00.200]period 1.3 billion accounts. That was in the same week they were talking to their
[00:27:06.280]shareholders about having 2.27 billion users worldwide. So when you see 1.3 billion
[00:27:12.460]accounts fake, that should tell us something strange. Are we really entering
[00:27:17.180]into what some people have been calling the dead internet? The dead internet is
[00:27:21.160]this idea that so much of the content, the users, are just not real. They're not
[00:27:25.640]human. And people, I can tell you, as we look at these conversations online, are
[00:27:29.800]commenting, even when they know the images aren't real, that that's kind of a
[00:27:34.420]depressing view of maybe where we're going. But I'll tell you what worries me
[00:27:37.900]most. Out of all the time that I spend, and I spend literally every day, seven days
[00:27:42.180]a week thinking about this kind of issue, the thing that worries me most is not
[00:27:46.640]that AI image that I showed you at the beginning of the talk. That's not what
[00:27:50.940]worries me most. That's concerning, but what worries me most is when there is
[00:27:55.740]actually going to be a bombing at the Pentagon or wherever, and people just
[00:28:01.740]think it's a fake image, and they don't respond. They become desensitized. They
[00:28:07.180]become jaded to the information environments in which we live, and we
[00:28:10.680]don't have them responding
[00:28:11.900]in ways that we want them to, and having the trust that we depend on in the
[00:28:16.160]society in which we live. That's what concerns me most. So, you know, generative
[00:28:21.920]AI, you know, in terms of the science, in sort of our world, in science, and
[00:28:26.840]research, and at university campuses, and in education, this of course is
[00:28:30.560]concerning. We have several research projects and papers that we're working
[00:28:33.500]on right now looking at the ways in which it's being used, of course, and
[00:28:36.500]misused, and also, you know, embarrassingly misused.
[00:28:41.620]We've been tracking some of these kinds of things. Some of you may have seen at
[00:28:46.200]least the most obvious example here. These are two papers, I have many others
[00:28:50.240]to show you, but I'm not going to, published by Elsevier. Elsevier is the
[00:28:55.080]largest scientific publisher, scholarly publisher in the world, and if you read
[00:29:00.920]the introduction you might be able to see it says, "Certainly, here is a possible
[00:29:04.340]introduction for your topic," and then on this other side it's, "I'm very sorry, I
[00:29:09.440]don't have access to real-time information
[00:29:11.340]or patient-specific data, I'm an AI language model."
[00:29:15.620]So it's one thing to use AI, I'm not fully against the use of these technologies to help
[00:29:20.840]us in the process, but we need to be more transparent, and holy cow, how did this get
[00:29:25.600]past reviewers, editors, and look at all the authors on both those papers.
[00:29:31.380]That's embarrassing.
[00:29:32.380]So that's the kind of thing, I mean, we're entering a world in which we're in this transition
[00:29:37.760]where this is what we're seeing and what
[00:29:41.060]the effects are are things that, of course, we're on the lookout for.
[00:29:44.280]One area in particular that I've been spending a lot of time on research-wise is looking
[00:29:48.080]at the ways in which AI is and potentially could engage on the peer review process.
[00:29:54.740]And I have all my concerns, of course, but an interesting paper came out from McFarland
[00:29:59.040]and others that looked at this massive spike in certain kinds of linguistic features in
[00:30:06.400]looking at a peer review corpus, this computer science corpus, after ChatGPT came out.
[00:30:10.780]This is fascinating to me, because if you look at these spikes and the types of words
[00:30:16.540]that are being used, you know, there still needs to be some work to determine whether
[00:30:20.640]that is just ChatGPT.
[00:30:22.980]But when you see that kind of change in the ways these peer review corpora are looking,
[00:30:29.160]that to me is of concern.
[00:30:31.360]Now, the other thing that's a concern, at least on the research side, I'm sort of moving
[00:30:34.820]away from the sort of more charged environments, the more charged political environments.
[00:30:40.500]I just want to talk about a general theme that's coming out in our research, too.
[00:30:44.620]I'm sort of hitting general themes that I've been seeing in our research, and I'm happy
[00:30:48.700]to jump into any of these research articles that I've talked about and research I've talked
[00:30:53.420]about throughout the talk during the question and answer.
[00:30:56.720]But there's this general brittleness that seems to exist in the research that me and
[00:31:01.820]my colleagues have been doing and others, many others have been doing as well.
[00:31:04.740]But let me just give you one example to demonstrate what I mean by brittleness.
[00:31:08.660]One is that there...
[00:31:10.220]You know, lack of physical understanding.
[00:31:12.500]Here's one of my favorite, it's weird to say older examples, like a year old.
[00:31:17.140]If you give this question, you gave this question to ChatGPT, 3.5, actually four does solve
[00:31:22.600]it a little bit better.
[00:31:23.600]It says, "I have a 12 liter jug and a six liter jug.
[00:31:26.720]I want to measure six liters.
[00:31:28.660]How do you do that?"
[00:31:29.660]It seems pretty simple.
[00:31:31.260]If you were human, what would you do if you wanted to measure six liters?
[00:31:34.640]You would use the six liter bucket, right?
[00:31:36.460]So here's what ChatGPT says, "Fill the six liter jug completely.
[00:31:39.940]Pour the water from the six liter jug into the 12 liter jug.
[00:31:42.080]Fill the six liter jug again.
[00:31:43.080]Carefully pour the water from the six liter jug."
[00:31:45.080]Okay, on and on and on.
[00:31:46.080]It clearly doesn't have access to the physical world, so we've been writing a set of papers
[00:31:51.700]that looks at the it's like an axiomatic paper that looks at what really what is
[00:31:56.370]it that these large language models are and what they're not they of course
[00:32:00.570]don't have any access to truth I know that seems obvious but that actually is
[00:32:04.190]important it doesn't have access to the physical world etc etc and we need to
[00:32:08.290]remind our students our faculty our researchers and the public that's
[00:32:11.850]engaging with this sort of thing about these axioms and so we have a couple
[00:32:15.570]papers out on on these axioms but let me give you an example now a more current
[00:32:20.090]example of this brittleness so if I was to give you this sequence here and if
[00:32:26.410]this is would be a typical analogy test that a cognitive psychologist would give
[00:32:30.050]and I would say something like okay here's your example ABCD goes to a BDC
[00:32:35.690]ABCD e now I'm going to give you a new example and I said I JKL goes to I JKL
[00:32:44.270]what
[00:32:46.590]and this is a test what would it be
[00:32:50.030]I JKL am all right good all right so you've just demonstrated to me as humans
[00:32:56.810]that you can solve this problem well there was a paper in nature human
[00:33:00.530]behavior that's been getting a ton of attention because it's pulling this new
[00:33:04.950]field called machine psychology which I'm a little skeptical of I'll just say
[00:33:09.590]from the very beginning this is published in some of the top journals in
[00:33:12.830]the world this has got so much attention it was talked about in all these
[00:33:15.650]different and where I'm going I'm having a back-and-forth literally even just
[00:33:19.010]yesterday with the authors about the methods that were used and I'm not
[00:33:21.890]nothing wrong with the authors asking these questions nothing wrong there it's
[00:33:24.790]just I question some of the the conclusions that are coming from the
[00:33:29.150]results of this paper in other words this is the results they show they said
[00:33:32.750]this they said okay look at this if you look at string analogies for example
[00:33:36.770]that's the ones I gave you if you look at humans that's in the lighter color
[00:33:41.210]and GPT-3 the darker color Wow look GPT-3 can solve letter string analogies
[00:33:48.050]Wow
[00:33:48.850]they must be smart like humans in fact now there's been studies testing whether
[00:33:53.830]they have theory of mind wait a minute theory of mind is being human knowing
[00:33:58.150]how another human thinks again just because they pass those tests doesn't
[00:34:02.530]mean they have these abilities what they're likely doing is memorizing the
[00:34:05.770]training data so my student and I decided to take this exact study and
[00:34:11.450]test whether they truly have what's called zero-shot reasoning that's the
[00:34:15.850]ability of solving something like letter string analogy
[00:34:18.790]without seeing any other examples that's what humans can do so we gave them
[00:34:24.070]synthetic alphabets we extended the sequences and long story short and you
[00:34:28.330]can read more about our response it's an archive and likely will be published
[00:34:32.770]later in nature about this that they fail miserably so again they're magical
[00:34:38.110]they're scary in terms of what AI is doing to our political conversations
[00:34:44.710]that's where I've talked about so far and they can make things like Sora a
[00:34:48.730]game amazing but they have a lot of weaknesses still and as we talk to our
[00:34:53.410]students about it we have to I think at least from my perspective we need to
[00:34:57.250]really engage in that but that's not stopping major governments from moving
[00:35:01.930]quickly forward with this technology I was recent I was recently talking to a
[00:35:06.150]Associated Press journalist and I was talking to this journalist about some of
[00:35:11.110]the things that we have been seeing in our research and the journalists had
[00:35:14.890]found him himself this is I'm quoted just
[00:35:18.670]as this came out a few days ago that New York City decided to just go forward
[00:35:23.470]with their AI chatbots giving advice to businesses and I guess to no surprise I
[00:35:30.850]guess I would you would think people would be surprised but I guess they're
[00:35:35.410]not these things were telling people how to break the law in other words so what
[00:35:41.290]happened was if you look at these questions you can't read them fully but
[00:35:44.890]it says like can I take a cut of my workers tips
[00:35:48.610]yes you can take a cut of your workers tips bosses can't take tips that's the
[00:35:52.810]reality you know landlords cannot discriminate by source of income but
[00:35:56.590]this chat bot was essentially giving advice that was breaking the law and
[00:36:01.730]here's the crazy thing New York City hasn't taken this chat bot down yet so
[00:36:06.910]we're moving so quickly into this space of like oh well we got to use it and of
[00:36:10.750]course in education we're worried about these things and and there is promise I
[00:36:15.130]don't want to be the person that wants to be the downer on
[00:36:18.550]everything it certainly can be used but this has severe consequences especially
[00:36:22.930]if it's in the medical field and we've got examples of those as well and by the
[00:36:27.310]way when it comes to elections these chat bots aren't doing well Princeton
[00:36:31.570]University came out with a recent study testing all of these different chat bots
[00:36:35.990]seeing if they could answer basic questions like where can you vote can
[00:36:40.030]you vote by text message and it says oh sure with a hundred percent confidence
[00:36:44.530]in the state of California you can text no you can't it's not true
[00:36:48.490]and yet people are using these engaging where we know that because of the usage
[00:36:52.090]of these chat bots and so there these are the things that we need to be aware
[00:36:56.590]of when it comes to the media literacy so there's this brittleness that exists
[00:36:59.950]but also of course legal issues and high consequential implications to moving
[00:37:08.010]quickly during this transition period so what the heck can we do about it this is
[00:37:12.270]of course the big question that I think we're gonna be talking about today and
[00:37:15.250]and campuses around the country and something that I'm
[00:37:18.430]really passionate about trying to engage in and this is sort of my
[00:37:22.310]reflection on it you know and my engagement on is that I think there's a
[00:37:26.290]little bit of technology we can throw at the problem I do think there is some
[00:37:29.710]policy actually I I've been working in the policy area and I'm not a policy
[00:37:34.330]expert I actually helped write one of the first deep fake bills in the state
[00:37:38.090]of Washington that passed and the governor you know was publicizing it and
[00:37:43.090]now I'm talking with people in DC about it and I don't think it's a perfect bill
[00:37:45.970]but at least it's you know a norm-setting
[00:37:48.370]kind of bill so there's some things we can do in policy there's something
[00:37:51.750]technology but really it has to be I think on the education side and I think
[00:37:56.850]that's what we do at a university obviously so I'll mention a few of these
[00:37:59.750]and others so a couple of my colleagues and I I'm on the board of this new
[00:38:04.270]nonprofit nonpartisan organization that we just actually was just featured in
[00:38:08.290]the New York Times a couple days ago called true media org and this is an
[00:38:12.130]attempt to bring the tools of detection of things like deep fakes which is not
[00:38:16.490]easy by the way and it's not perfect
[00:38:18.310]but we want these tools to be available to the public and to organizations that
[00:38:23.410]don't have large budgets like tech companies or the Democratic National
[00:38:27.070]Convention or the committee or the Republican National Committee we want
[00:38:30.910]everyone to have access to these during elections and so we've started this
[00:38:34.250]nonprofit and that's an example of kind of the technology I think we can throw
[00:38:37.750]at it but I know from even using the best things that we know of that it's
[00:38:42.430]not they're not they're not perfect like I
[00:38:45.430]said here's the bill that passed in the state of Washington there's some things
[00:38:48.250]around deepfakes going in different states that we can do and I don't think
[00:38:51.610]they're perfect bills but again the way I look at this exercise is just a flag a
[00:38:56.930]norm-setting flag that says okay there needs to be some friction in the use of
[00:39:02.010]these technology around major consequential events like political
[00:39:05.310]campaigns now there's major decisions going on in the Supreme Court for those
[00:39:08.950]that follow things that are going on in the Supreme Court and certainly this
[00:39:12.790]comes close to researchers and anyone working on misinformation and in the
[00:39:15.970]social media space this
[00:39:18.190]is one of the biggest cases we'll see in this year there's several others but
[00:39:21.610]this is the big one Murphy versus Missouri this is this idea or this
[00:39:25.750]question of whether the government and the White House in particular where they
[00:39:28.910]can have conversations with tech companies and after the this already
[00:39:33.370]played out a couple weeks ago the decision is coming in a couple months it
[00:39:36.190]looks like most of the justices are going to say that it is okay to have
[00:39:41.890]conversations between government and one of the things I think that convinced
[00:39:45.030]them is one of the arguments that was made was well
[00:39:48.130]okay justices and this happened to one of the justices if someone was to get
[00:39:53.150]access to your personal information and was to docks you're gonna make and we
[00:39:58.930]knew that was going to be made available can we go tell the tech companies to be
[00:40:02.530]aware of that they're like oh yeah actually we probably don't want our
[00:40:05.570]personal information being so those are the kinds of things that the government
[00:40:09.310]maybe should have access but that's still being debated I think there should
[00:40:12.070]be limits I don't think they should be core coursing tech companies but this is
[00:40:16.030]an interesting one the most successful one that's
[00:40:18.070]going on that's bipartisan support is the Defiance Act so we have a couple
[00:40:21.970]projects looking at the spread of non-consensual pornographic images and
[00:40:28.210]this is something that both Republicans and Democrats think that we should act
[00:40:32.170]on pretty quickly and I hope it becomes a model for how we can do something in
[00:40:37.270]the technology space okay so that was a little bit of tech that was a little
[00:40:40.930]policy and I'm gonna end the talk here talking about those programs so I
[00:40:43.790]mentioned this miss info day it's something I'd love to do in all states
[00:40:46.630]if you want it we have make everything
[00:40:48.010]free we actually just had our program this is something that we make all the
[00:40:51.950]tools available we make that available to the teachers the universities the
[00:40:55.450]librarians let us know if you're interested my favorite program that
[00:40:58.870]spurred from that is something where teachers now are teaching high school
[00:41:03.070]students for an entire semester about miss around media literacy and then they
[00:41:07.390]teach their grandparents their neighbors and their their parents and friends at
[00:41:12.590]an event and I've been to a couple of these events they're amazing and we call
[00:41:15.770]these sort of media ambassador nights not
[00:41:17.950]because of the political charge I spent a lot of my time I love games so I've
[00:41:22.390]been making games many years ago when style Gans and deep fakes were just
[00:41:26.190]coming online I created a game called which faces real calm it's been prayed
[00:41:29.950]tens of millions of times around the world in different languages just to
[00:41:33.130]bring to public attention that the technology exists because it's during
[00:41:36.490]that transition time that I think we're most vulnerable and so I just the games
[00:41:41.190]easy one image is real one is not you have to decide and then of course you
[00:41:46.130]know sometimes you get it right sometimes you get it wrong
[00:41:47.890]I'll let you decide which one's real we may go back to that we've also now
[00:41:52.150]created some escape rooms that we can deliver to you we send them if you want
[00:41:55.750]them again free we do a virtual one we do digital ones so and we've translated
[00:42:00.190]these in different languages it's an escape room for students adults
[00:42:03.130]communities we've created lots of versions of it and the idea is to engage
[00:42:07.510]in those things so all these resources and more I'll put this link up if you're
[00:42:10.570]interested and then I'll just say that also we think that coming offline is
[00:42:14.890]really really of course how we can address some of this
[00:42:17.830]so we have a program funded through the National Science Foundation that we
[00:42:21.430]call co-designing for trust where we work with communities rural librarians
[00:42:25.450]community colleges high school educators and we and community leaders to design
[00:42:33.150]interventions to misinformation as I mentioned I have you know books and
[00:42:37.570]things around this and this is what we're dealing with in the education is
[00:42:41.110]what are we gonna do with AI in education and by the way don't do what
[00:42:44.050]Boston University did which was say okay we're gonna
[00:42:47.770]strike and we're gonna replace you with AI that was kind of a problem so I know
[00:42:53.290]that we're ending so I'm gonna skip a few slides here so and I will just say
[00:42:58.110]that also there's a really interesting things going on on the internet around
[00:43:02.890]the amalgamation of traditional search and faster search and I'll just say
[00:43:08.470]there's so many of these examples I have but I have to show one that got a lot of
[00:43:11.950]attention was written in the Atlantic Monthly about it but if you ask a
[00:43:15.310]question like country in Africa that starts with
[00:43:17.710]K while there are 54 recognized countries in Africa none of them begin
[00:43:24.170]with the letter of the reason why that happened was that this was an AI
[00:43:40.470]generated website that's just creating garbage and Google goes oh wonderful
[00:43:47.650]at nine jumping it now so the last thing I'll say and I'll end here is that
[00:43:53.950]something that we hold it for though is that companies like perplexity where they
[00:43:59.590]say if you can directly answer somebody's answer nobody needs those two
[00:44:03.110]ten blinks any links anymore I say that is exactly what we don't need in this
[00:44:08.650]new era we absolutely need those links we need to be able to do what these three
[00:44:13.270]dots allow those are the three dots that allow you now which we encourage Google
[00:44:17.590]to do and they now allow you to look at the metadata so with that I'll end and
[00:44:21.010]say thank you so much sorry going over a little bit but happy to take any
[00:44:25.090]questions
[00:44:27.470]give any questions sorry I know I went by really fast and I know we're getting
[00:44:41.710]caught but happy to talk with others oh yes
[00:44:47.530]you were the one that we're feeling we're gonna be screwed so we're on the
[00:44:51.070]same page probably no no I'm fascinated by the the misinformation and hamp a
[00:44:54.910]schoolers and was just curious I mean I know Washington is a little more
[00:44:58.450]progressive than some other places but you know have you gotten any pushback
[00:45:01.210]from that as well I'm just thinking about here in Nebraska if you were to
[00:45:04.390]bring something like that here you probably have some people saying great
[00:45:06.930]and you'd have other people saying you know it's a liberal conspiracy yes no
[00:45:10.810]I'm glad you brought that up so the state of Washington is about as red as
[00:45:17.470]maybe more red maybe even Nebraska in some ways and very very very blue and so
[00:45:22.390]we a lot of the work that we're doing are in those rural communities in
[00:45:26.370]eastern Washington we're working with Washington State University and some of
[00:45:29.450]the other universities that have better access to the rural communities and the
[00:45:33.150]one thing we learned a lot of things first of all don't call it Miss Info
[00:45:35.710]Day that's the one thing we couldn't do that so we call it media ambassadors and
[00:45:39.790]actually it's one of the things I talked to and I grew up in a very small
[00:45:42.790]community so I enjoy going to those communities and talking with them and I
[00:45:46.390]talked to him about how
[00:45:47.410]one of the most important skills that we can give our kids today it's like a
[00:45:50.890]superpower I was always I love superheroes when I was young Superman
[00:45:54.170]Batman all those things I said actually the superpower today is being able to
[00:45:57.430]discern what's true or not
[00:45:58.870]don't you want your kids to have that ability and we're very careful in the
[00:46:02.350]content we use we we try to customize it to the things that that community cares
[00:46:06.850]about so if it's an agricultural community we talk about agricultural
[00:46:09.790]issues whether it's GMOs or other things that that people care about we try if we
[00:46:14.410]do move into the political space which we don't very often
[00:46:17.350]we really try to give examples because it really does exist on both sides and
[00:46:21.970]then what we do is is is really an actual a lot of examples we give of
[00:46:27.010]other countries too it's easier to talk about these issues when it's that far
[00:46:30.490]away and that's like well of course that's a silly thing I can't believe
[00:46:33.310]that that's what happens and you go oh my gosh there's a parallel hill going on
[00:46:36.230]but it's very it is very difficult I won't say it's not easy it's that it's
[00:46:39.250]easy it's very very challenging but I I feel like it's the only way we have to
[00:46:43.670]do it in both sides and so that's why we do the co-designing we go in and
[00:46:47.290]say you tell us you know what what is it that you want your your kids and your
[00:46:52.150]community to know so it's an ongoing project and a hard one wonderful
[00:47:00.490]wonderful presentation and I appreciate your emphasis on education on media
[00:47:06.310]literacy actually this afternoon I'm about to present a counter case on that
[00:47:10.530]okay yes okay so um I'm curious about the Miss info day and related education
[00:47:17.230]oh modules from my own research we tend to find that such modules are effective
[00:47:23.710]in the short term but they don't necessarily stick for the long term and
[00:47:28.030]sometimes they bring backfire effects that is they actually increase overall
[00:47:32.830]skepticism not only towards misinformation but towards trustworthy
[00:47:37.290]information as well I'm wondering if you couldn't speak about that thank you yeah
[00:47:41.350]so I think this is a really really important issue and I'll say let me just
[00:47:47.170]kind of summarize about eight years of like thinking about this and and you
[00:47:54.790]know in doing research and developing teaching modules in this one of the
[00:47:57.770]things that I've been trying to convey just in the last year so and this is
[00:48:00.910]informed my colleagues that I that I talked with about that during the
[00:48:06.050]pandemic there so much of the world had to do their own research because there
[00:48:10.130]was nothing there and doing your own worst research is something I highly
[00:48:13.610]encouraged but
[00:48:17.110]if you come back to me and you say I found evidence to support some claim
[00:48:22.090]then I think that's not there's infinite evidence on the internet now what you I
[00:48:29.730]want you to do is come back to me with how you came to your evidence tell me
[00:48:34.990]how you would actually separate one piece of evidence from another and if we
[00:48:38.470]can change that culturally that's something that I think is one of more
[00:48:43.030]important things to do and so I bring that up because when I have conversations
[00:48:47.050]about like epistemology which I never use the term when I talk to public about
[00:48:50.170]just like how do we know what we know and we've done these like you know
[00:48:53.650]epistemology in the real world sort of conventions you know with the public
[00:48:58.090]that I see you know just anecdotally less backfire resistance to these kinds
[00:49:04.390]of things and actually as you know because you read about this there's
[00:49:07.870]there there's evidence that shows backfire like when Brandon Nyhan and
[00:49:11.770]others actually came out with this even he's come back and said actually the
[00:49:14.590]backfire effect might not have as much of an
[00:49:16.990]effect I mean this is all debated and now you know maybe you're going to even
[00:49:20.170]get into this but at least just from my personal experience when I change the
[00:49:24.730]conversation from just evidence or you know what's a reliable you know what's a
[00:49:31.210]reliable news site versus another and I just talk about how you go about doing
[00:49:35.350]that and have them reflect on that I find that that's the least
[00:49:38.890]I see the least resistance in that sort of thing and then when I talk about
[00:49:46.930]you know just generally misinformation to political leaders and business
[00:49:51.190]leaders I find that one area in which I can make a little bit of progress is
[00:49:55.870]when we talk specifically about the crisis element rather than just talking
[00:50:01.210]about you know it's easy to talk about the political environment because that's
[00:50:04.510]the one that we have the you know the most visceral the strongest visual
[00:50:07.630]responses to um but yeah I I still think it's something
[00:50:11.350]we we can't ignore because I would hate to be doing all this work and then
[00:50:16.870]having people become more skeptical in fact generally I actually am concerned
[00:50:22.510]about even talking about sometimes misinformation to the public and about
[00:50:26.050]the ways in which our information environments aren't all that reliable
[00:50:28.810]because oh my gosh I don't want them leaving like well I don't care like I
[00:50:32.470]give talks for example on misinformation in science which I am a huge fan of
[00:50:37.090]misinformation science and I've written many papers about um how much you know
[00:50:41.050]it still works despite these problems
[00:50:43.030]um but I do feel like there probably has been some talks that I've given
[00:50:46.810]or other things where people leave even more concerned more skeptical more
[00:50:51.430]cynical the one thing that really stuck with me is when I talked to a middle
[00:50:54.670]school class several years ago actually this was before the pandemic and I
[00:51:01.030]showed them some pieces of information I said you know you know which one do you
[00:51:04.150]think is more reliable or not and and a couple of students that raised their
[00:51:07.210]hand they said well none of it's true I don't trust anything on the internet and
[00:51:11.350]I don't want that to be the case I think that's a worse output for all the things
[00:51:16.750]so if that's where we're going I'm just gonna stop go get a piece of ice cream
[00:51:20.110]and call it a day because I don't want to make people more skeptical and I
[00:51:23.990]think that's probably what you're getting at so yeah thanks for the
[00:51:26.350]question I mean it's really important yes thanks for such a wonderful
[00:51:34.690]presentation and it's very resourceful indeed so your title site said that
[00:51:40.990]misinformation is already challenging and then came generative AI so my question
[00:51:46.690]is how severe do you think the misinformation was challenging before
[00:51:52.330]the generative AI and how much do you think generative AI has magnified it and
[00:51:59.350]what are the areas that you think misinformation was challenging before
[00:52:03.850]generative AI came into the picture no it's a great question and it's a good
[00:52:07.810]chance to do that comparison when you work in technology studies history is
[00:52:11.630]yesterday you know and you're like you study the future really
[00:52:16.630]so it's just things are moving so quickly that we can do those comparisons
[00:52:20.830]and so in the misinformation was a challenge before I guess if before was
[00:52:26.990]the sort of rise of social media social media all by itself and the ways in
[00:52:30.830]which it would sequester information it would amplify information and and because
[00:52:36.830]of this connectivity that was hyper connectivity that we have you know
[00:52:41.050]before we didn't we didn't even need generative AI to cause challenges here and also
[00:52:46.570]so you have all these different forces so one you have algorithmic sorry you
[00:52:52.090]got this I have this ring I'll step back so you have this algorithmic element
[00:52:57.370]that's you know as Kirsten talked about it's it's sort of driving you know a lot
[00:53:04.810]of the issues that we're seeing here in terms of even just our own personal view
[00:53:09.370]of the the world and also you have bad state actors we've seen that in our
[00:53:13.870]research actually some of my colleagues work on this every day
[00:53:16.510]that happened and they were effective before but the difference there is that
[00:53:20.410]you had you could write bots but you couldn't write them as effectively and
[00:53:25.570]create the sort of plausible writing output that you can now with generative
[00:53:29.450]AI and and the other thing that that that we saw at least when we compare
[00:53:32.950]like the 2016 election to the 2020 election a lot of what we saw that was
[00:53:37.810]you know a lot of what was driving some of the issues the sort of macro level
[00:53:41.530]issues were inauthentic accounts outside the
[00:53:46.450]United States in the 2020 election we saw more of what's called we call
[00:53:51.490]participatory disinformation that's when people are given a frame so here's
[00:53:56.650]a frame a narrative of some sort you find the evidence so then people go
[00:54:01.490]looking for the evidence they drop things into that frame and then it
[00:54:05.590]amplifies you don't need generative AI to do that you just need to create the
[00:54:08.570]frame once you create that narrative that frame whether it's you know a stop
[00:54:13.430]to steal or whatever it is you have that and then people look at and a lot of
[00:54:16.390]sometimes that evidence is legitimate and you should look into that but it but
[00:54:21.070]it creates that you don't need generative AI for that now what's
[00:54:24.530]happened of course now is that we see experimentation going on at a scale like
[00:54:31.090]that you can only do with automated bots that can create plausibly looking
[00:54:36.370]language and all sorts of other modalities although I'll say that the
[00:54:40.270]modality that scares me the most right now is probably audio that's the hardest
[00:54:44.110]to discern I think for us and it's
[00:54:46.330]now down to 15 seconds if you give me 15 seconds of your bit your voice your
[00:54:51.010]question if I just could have recorded your question that you asked me I can
[00:54:54.670]create a deep fake of you saying something preposterous and so that to me
[00:54:59.590]is there we saw that with robocalls in New Hampshire that was kind of a low
[00:55:02.930]level sort of application of it but I think these these are challenging it's
[00:55:07.430]just that what happens is we start to develop the understanding and tools
[00:55:11.530]about what to do about the spread of misinformation social media and then all
[00:55:15.430]of a sudden generative AI
[00:55:16.270]came and now we're trying to generate the tools and understanding of what's
[00:55:19.450]going on and it's it's really chat it's especially challenging but I would say
[00:55:22.970]the thing that's changed the most from the world in which social media was
[00:55:27.190]amplifying things and the biggest challenges we had there were these
[00:55:31.870]little isolated islands of communities that we couldn't we can't track
[00:55:34.930]everything and when we saw like whatsapp emerge in places that have never had the
[00:55:39.490]internet in in places in the world that didn't have internet for the first time
[00:55:43.730]even Edgar Welch who was the person that went shooting on
[00:55:46.210]pizzagate in Washington DC he was someone that had just had access to the
[00:55:50.730]internet to the most part and it's it's that transition into the information
[00:55:55.510]worlds that's always going to be a problem but the new thing with
[00:55:58.050]generative AI I think really is experimentation the scale and the
[00:56:01.630]plausibility of what we see yes so let's jump to a worst-case scenario okay let's
[00:56:10.930]do that let's get really screwed is Cambridge Analytica 2024 coming for us so
[00:56:16.150]let's take generative AI and sophisticated prompt engineering and a
[00:56:21.850]Facebook network of 1.3 billion fake accounts what could go wrong what are
[00:56:28.030]you afraid of so I absolutely think it's a possibility and it's likely playing
[00:56:33.790]out right now I I wouldn't you know it's we're we're trying to you know keep track
[00:56:41.710]of these kinds of things the things the thing about Cambridge Analytica
[00:56:46.090]is they were basically leveraging tools that anyone could have had access to
[00:56:50.650]they just leveraged them you know at the right time and had the right resources
[00:56:55.570]at the time to be able to do that that's it's clearly happening now and I think
[00:57:01.750]here's here's getting back to the previous question to the the thing that
[00:57:06.550]I find most challenging in dealing with this now is that we'll tell the public
[00:57:11.290]we'll say okay if you see something strange do
[00:57:16.030]what's called you know lateral reading so to open up a new tab look at other
[00:57:19.250]sources but what's happening now is that their technology allows you to fill all
[00:57:24.850]those extra tabs with other news sites and even comments on new sites and other
[00:57:30.790]things that and that can be customized if cookies are being saved and so what
[00:57:34.590]would Cameron Cambridge Analytica it was on Facebook but now it's not just
[00:57:39.730]Facebook in which you can sort of do this kind of experimentation and and do
[00:57:44.230]these you know and apply some of these cycle
[00:57:45.970]metric measures I mean some people are sort of skeptical of what they can
[00:57:49.030]actually do but yeah I absolutely think we cannot rule out the fact that we're
[00:57:55.950]constantly being barraged with a lot of accounts that aren't even really humans
[00:58:01.090]and that are running experiments constantly so I can give you a couple
[00:58:05.030]papers that came out of the Stanford Internet Observatory that's addressed
[00:58:08.710]this exact issue but yes I what could go wrong lots and it does keep me up at
[00:58:15.010]night for sure
[00:58:15.910]I saw a question here yes you mentioned earlier like what is news and your I
[00:58:25.570]can't help but think like that
[00:58:26.950]I don't know how to articulate this question but you keep talking about
[00:58:29.230]systems of power and you don't want to talk about politics and you don't want
[00:58:32.510]to talk about it at what point is this systematic and and and what is this
[00:58:37.630]really about AI because when I see true media.org I wonder who owns it and I
[00:58:43.030]wonder who owns you and I wonder why
[00:58:45.850]you're telling me the thing you're telling me and and I think that like
[00:58:48.550]everybody has to critically ask of everything that but but Cambridge
[00:58:54.310]Analytica happened before AI the the media powers that are out there still
[00:58:59.290]want to misinform like the only thing the internet wants to do is sell me they
[00:59:03.150]don't there's no information there's no altruistic this is my perspective at
[00:59:06.470]this point but there's no altruistic platform out there that wants to unite
[00:59:10.850]humanity the internet is just to sell you know like so I think and I'm just
[00:59:15.790]channeling some of my own interior but how do you differentiate between what
[00:59:20.110]where this problem is systematic and where it's human impulse to trust the
[00:59:25.990]three sources or to educate myself and do my own research because if we're all
[00:59:32.070]doing our own research as critically and analytically as possible inside of a
[00:59:35.750]platform that really doesn't want us informed we're never gonna get there and
[00:59:39.210]I'm just I'm curious about like where does misinformation stop if we tackle AI
[00:59:45.730]to me misinformation is about questioning the source not the actual
[00:59:50.070]content but it's what's the incentive of this person to tell me the truth and how
[00:59:55.870]do you would I guess how do you see that yeah I'm so glad you brought that up
[01:00:00.010]because it's one thing I had to blow by real quick so one of the things I wanted
[01:00:03.550]to end with was say that one of the best things we can do is check the source so
[01:00:07.570]there's this method that my colleague Mike Caulfield came up with it's called
[01:00:10.150]sift basically it's real simple it's stop investigate essentially the source
[01:00:13.930]that's one of the most important things you could
[01:00:15.670]do of all the things it's not the content itself it's really the sourcing
[01:00:18.910]and credibility and reputations matter more in this world than it ever is and
[01:00:23.110]you should we should be questioning it to some degree at some point we have to
[01:00:26.530]have some trust I mean trust levels have gone down in about every institution
[01:00:30.610]possible but to give you a practical example not to say this is a panacea for
[01:00:35.050]addressing this issues and it's certainly more than AI I mean we had
[01:00:38.410]this issue if AI wouldn't have come into that I mean first of all a I's been
[01:00:42.050]around of course for you know very very long time many many decades since
[01:00:45.610]Rosenblatt in the 1950s I mean this is not a new thing but its current form
[01:00:51.310]where you have these machines that can create plausibly looking text that is
[01:00:55.630]new but it's not like misinformation is I completely I could not agree more that
[01:01:00.990]misinformation is is its own thing and it didn't need AI to make it worse but
[01:01:06.050]there are some practical things we do so one of the things we've been telling
[01:01:08.710]Google for like three years maybe four years now is that why aren't you making
[01:01:13.270]the tools that fact-checkers need all the
[01:01:15.550]to do sourcing so this is sourcing like you say so you can check what you know
[01:01:20.590]here what this means like back to bar it was just recently in Finland talking to
[01:01:24.610]this fact-checking organization I want to know who they were so you can now go
[01:01:27.810]to these three vertical dots they exist on Google now not that they're perfect
[01:01:30.710]Google is now providing some of that fake data we encourage them to do this
[01:01:33.990]and now they are doing this and you can actually see you can say you know it was
[01:01:38.690]first indexed by Google in March 2014 if you see something indexed yesterday that
[01:01:42.450]might be something worth question there's all this metadata again this is
[01:01:45.490]insane oh we fixed it it means we need to create better tools for sourcing and
[01:01:49.750]not doing what perplexity asked us to do which is just just accept the answer
[01:01:53.350]we're gonna give you the answer you don't need those ten blue links anymore
[01:01:56.250]and so the example I love to give because this is something I study I can
[01:02:00.330]guarantee you every like in the next hurricane we have there's gonna be shark
[01:02:05.050]swimming in the street guaranteed and it's gonna be this one now this the kind
[01:02:09.430]of the crazy thing is last year there actually was a shark swimming in the
[01:02:13.550]street so now my students make fun of me again with this
[01:02:15.430]sample but still a good example but with these three vertical dots this
[01:02:18.450]literally is a new feature within the last several months you can now look at
[01:02:22.790]the three vertical dots of images and you'll see with this image right here
[01:02:25.990]that Snopes and others have said does this picture show a shark no and you can
[01:02:29.810]look at all these places now it could still go back to your point and say
[01:02:34.030]well ultimately and I don't know if that's you know I still don't know if
[01:02:38.870]that's true and that's why we have to get out of that internet selling
[01:02:42.130]business and come in line like this come into our communities I
[01:02:45.370]think we need to create a culture of coming offline more than anything so that
[01:02:49.150]we can get back to that kind of thing
[01:02:53.130]it totally is it we're coming more splintered in fact the term that I me
[01:03:02.410]and other colleagues have been using for a long time it's not my term but I use
[01:03:04.810]it it's called the splinter net we know it's not the internet anymore in the
[01:03:07.690]1990s I remember sitting around with my friends going wow democratization of
[01:03:11.770]information it's gonna be so cool you know sitting around the campfire now
[01:03:15.310]like it's just on the opposite direction like I don't think that's where we've
[01:03:19.210]gone we have our own internets now and if we get away from it at least we share
[01:03:23.030]a physical space but anyway that would be a little bit over simplistic with the
[01:03:28.090]problem but to your point yes it is more than just AI misinformation does involve
[01:03:33.290]politics power and all those things that have been around forever so thank you
[01:03:38.930]thank you Jenna

The screen size you are trying to search captions on is too small!

You can always jump over to MediaHub and check it out there.

Comments

0 Comments

Misinformation was Already Challenging — Then Came Generative AI

Description

Searchable Transcript

Comments icon comment

Related Channels

Comments