Journée d'études 2025 de l'Arcom : La vérité en crise : politiques et technologies de vérification

Arcom

https://www.arcom.fr/actualites/quatrieme-journee-detudes-de-larcom-presentation-des-travaux-des-chercheurs-sur-les-medias-audiovisuels-et-numeriques

Transcript

00:00:00Donc, le panel suivant va commencer.

00:00:04Il s'inscrit dans une parfaite continuité avec ce que vient de nous présenter Thomas

00:00:09Hochman, puisque ce panel est consacré à différentes facettes de la désinformation,

00:00:14qu'il s'agisse de l'information de l'opinion, de ses conséquences sur le processus électoraux,

00:00:18ou encore sur les technologies dont elle peut se nourrir.

00:00:22Donc c'est une session qui est entièrement en anglais, donc je vais changer de langue

00:00:25si vous le voulez bien.

00:00:27Donc, nous sommes délicés à recevoir un groupe de chercheurs qui couvrent un grand nombre

00:00:33de fields, de communication studies, de l'économie, de l'économie, de l'économie, de l'économie,

00:00:38de l'économie, de l'économie, et de l'économie, de l'économie, de l'économie et de l'économie.

00:00:44Dans la prochaine hale, nous allons parler d'une variée de méthodologiques qui rendent ce panel,

00:00:50je pense, particulièrement excitant.

00:00:52Donc, nous allons commencer avec Alisson Preston à l'économie.

00:00:55Bienvenue, Alisson.

00:00:56Donc, Alisson, nous allons parler de pourquoi les gens ont des vies que peuvent se déferir

00:01:04de la vue du public, comment ces beliefs se font, et ce qui permet l'économie de l'économie

00:01:11de l'économie et de l'économie.

00:01:13Alisson, vous êtes le chef de l'économie, de l'économie, de l'économie et de l'économie.

00:01:18L'économie, l'économie, l'économie, l'économie, l'économie.

00:01:23Si vous ne connaissez pas, c'est le régulier de l'Économie,

00:01:27dans le domaine de l'économie, donc TV, la radio, la radio, et donc on.

00:01:30Donc, c'est le régulier de l'économie, que l'économie,

00:01:34donc, ça fait plus de plus de plus de sens pour l'économie avec vous, Alisson.

00:01:39Et votre travail va nous aider à comprendre

00:01:43comment les gens pour les résiliers de la vie des narratifs.

00:01:46Alors, la parole est yours.

00:01:48Super, merci, et je suis désolée, mais je vais parler en anglais.

00:01:54So, as Olivier says, I'm Alison Preston,

00:01:57and I'm Head of Media Literacy Research at Ofcom.

00:02:02Thank you so much for inviting me today.

00:02:05It's very good to be here.

00:02:07And I'm going to be talking about people's attitudes to news information and misinformation.

00:02:16What are they thinking and feeling about what they are consuming?

00:02:21But I'm going to start with some context about the current ways that people understand news practices

00:02:30and production norms and their broad attitudes to news.

00:02:35So, firstly, I do want to hear from you.

00:02:46Let me just go back.

00:02:47I don't know whether the builds on this slide work.

00:02:51Maybe they don't.

00:02:54Maybe they don't.

00:02:56Never mind.

00:02:57But I am going to get a show of hands.

00:02:59So, ignore the fact that you've just seen that.

00:03:01So, who here in the room believes that the more a story is edited,

00:03:06the less it's likely to be true?

00:03:08So, the more a story is edited, the less it's likely to be true.

00:03:13And I can't see any hands on that one.

00:03:17So, who believes that, who agrees with the statement,

00:03:21I trust eyewitness news video footage more than photos or written articles?

00:03:28So, who trusts eyewitness video more than photos or written?

00:03:34There's one person there.

00:03:37Couple.

00:03:38Not many.

00:03:39And then, who believes or agrees with the statement,

00:03:42journalists usually follow codes of practice to encourage accurate and ethical content?

00:03:49So, who believes that journalists usually follow codes of practice to encourage accurate and ethical content?

00:03:57And yes, many of you do agree with that statement.

00:04:01So, there we go.

00:04:04That is, you are very different from the UK population as a whole.

00:04:10Because as you can see there, considerable numbers of people believe those first two statements.

00:04:18There are differences by demographic group, by audience groups.

00:04:23So, men, younger people, and people from minority ethnic groups are more likely to agree with those top two statements.

00:04:31So, those findings are pretty stark.

00:04:34They do show that there is uncertainty about what to trust.

00:04:39And as you can see here, less than half of UK adults, 45% there, feel confident judging whether or not sources of information are likely to be truthful or not.

00:04:52And again, younger people, and people from minority ethnic groups and men are more likely to feel confident about those things, as well as people from higher socioeconomic groups.

00:05:06So, these views and this kind of type of confidence, I think, provide useful context for understanding the extent of disaffection or otherwise with mainstream or traditional news providers.

00:05:22And this is a question that we ask, that I'm about to show you, that we ask on our media literacy tracker, a longitudinal survey.

00:05:32And we ask respondents to decide which of the following statements are most true to them.

00:05:39So, this one is a bill, that's great.

00:05:42So, first of all, 16% of people agree with the statement that, you know, they use mainstream news media and they trust it.

00:05:53And then four in ten are what we might call media literate, that they use mainstream news sources, but they sometimes question whether or not that information is accurate.

00:06:05One in five are what you might call sceptical.

00:06:10They use those sources, but they always question whether the information is accurate or reliable.

00:06:17And then six, don't really think about it.

00:06:20And then 6% are rejecters and 10% are sort of disengaged.

00:06:27They're not using mainstream news for various other reasons.

00:06:32And on the right there, you can see that that combination of rejection and disengagement rises to a quarter of people from a DE, socioeconomic background.

00:06:43So, that's people with semi or unskilled jobs or those who are unemployed.

00:06:49So, I think this is one aspect of this kind of contextual picture and getting understanding around attitudes towards misinformation.

00:06:59I.e., significant minorities of people are feeling ambivalent about some of the basics of professional news and professional news production.

00:07:09So, in terms of where people are getting their information on in order to make up their minds, in order to decide what they feel is correct, we ask, again, another question that focuses on whether or not people prefer to use the opinions of those around them or, on the other hand, news providers or official sources in making up their minds around the news.

00:07:37And if they felt that both of these things are useful, that's absolutely fine.

00:07:43They kind of sit in the middle.

00:07:45And so, you can see here that, indeed, 4 in 10, you know, not really sure whether or not it's people around them or news providers and other sources.

00:07:57But, as you can see there, 1 in 5 are mostly reliant or prefer to rely on people that are kind of around them or more informal sources.

00:08:08There is a difference by age, and so, again, as you can see there in pink, those aged 18 to 24 are considerably more likely to use those kind of local friends and family networks.

00:08:28So, our research is illustrating that people are judging information and whether it can be trusted by using their own worldviews, their own knowledge base, and their own beliefs as bellwethers.

00:08:43And, of course, this is nothing new because we all use our own views and experience to decide whether or not something is trustworthy or believable.

00:08:54We don't encounter news with a kind of neutral rationality every single time, but with a set of existing beliefs and opinions about that situation.

00:09:04But what's different is how those opinions are being formed.

00:09:09And so, as you can see, for some, it's very much happening away from mainstream or official sources and by being more reliant on these more informal sources of news.

00:09:21So, what we've got is a picture of fluid news consumption patterns, less reliance on news organisations, and less belief in them.

00:09:30So, that's the context within which misinformation narratives or minority beliefs can develop.

00:09:40So, moving on to our qualitative work, we've carried out two major pieces of qualitative work in recent years.

00:09:55One was from two years ago and one was from earlier this year.

00:09:59So, the first one was to understand why people hold the beliefs they have.

00:10:04And then, more recently, we've talked to people that have moved away from such beliefs to understand their rationales and what's kind of catalyzed that change of mind.

00:10:15And we also wanted to explore in that work as well really kind of what works in terms of changing minds or what works in terms of the messaging that might be most helpful.

00:10:26So, in that first research, we used the term minority beliefs throughout the research and then indeed in the publication as well.

00:10:38What we meant by that was views that are not widely held by the general population.

00:10:43And we used that term as a kind of neutral and non-judgmental way to enable those potential participants or indeed the participants in that study to feel more comfortable about identifying their own beliefs and talking about them.

00:11:00The topics that we focused on were, as you can see there, health protection, climate change, and the Russia-Ukraine conflict.

00:11:11And we wanted to get a sample of people that, you know, believed things about each of those issues.

00:11:18And so, what we wanted to understand was how these beliefs are formed, how they change and so on over time, and also the intersection with media literacy.

00:11:28For the second piece of research, we carried out, first of all, a set of interviews with a number of people, including 18 that had previously held particular beliefs but no longer did so in order to see, you know, what were the catalysts for change.

00:11:51And then we also carried out workshops with around 65 people in total to explore what they felt were the most useful ways of countering mis- and disinformation narratives.

00:12:08So, focusing on that first piece of qualitative research to begin with, I think one of the things that was interesting was that while you might think,

00:12:21and while I think it's often can be characterized that people with certain types of minority belief can be seen as very dogmatic and fixed in their view,

00:12:32what we found from this research was that they felt that they really weren't,

00:12:38that they felt that they needed quite a bit of personal resilience to hold on to the beliefs that they had,

00:12:46but at the same time they didn't necessarily feel those beliefs were clear or fixed.

00:12:52And so what they were doing instead was to, they felt that they were questioning,

00:12:59that they were questioning dominant narratives, and that they were being active and inquiring about what those really meant,

00:13:09and delving into them in more detail and so on.

00:13:11So, again, as you can see here, they almost kind of characterized themselves as being quite proactive researchers,

00:13:19that they wanted to uncover what was really going on and so on.

00:13:24And they found, you know, they would talk of others that followed mainstream news or, you know,

00:13:31believed mainstream information without doing the critical inquiry as being very kind of passive and fixed in those own beliefs.

00:13:39So, I think that, you know, one of the things that was coming out of this, you know,

00:13:48was the way that people said that they would be talking to others,

00:13:52that they would be researching online, that they would be exploring and comparing and contrasting news sources

00:13:58and all these sorts of things, which, you know, as you will doubtless agree,

00:14:05many of those things could be described as being fairly media literate responses to their kind of quest for information.

00:14:14But I think what is interesting is this, you know, what happened was that there was very,

00:14:25there was a kind of variability in how this media literacy or this kind of type of media literacy was applied.

00:14:33So, what tended to happen, what people tended to say was that they were giving far more credence to that kind of video testimony,

00:14:43apparent first-hand accounts and so on, that those were being seen as particularly reliable

00:14:50and not the evidence that might have been mediated through other types of actor or journalist and so on.

00:15:01And one of the things that was also striking was the way that they would often describe how the absence of information

00:15:10or the lack of representation of certain either kind of videos or points of view on the mainstream news

00:15:18was in and of itself evidence that that information was being suppressed or those views were censored.

00:15:26So, they didn't kind of cognitively think, well, actually, it might be because it's not, you know, not reliable.

00:15:31It was, no, that's being suppressed and that's problematic.

00:15:36So, one particular example that sticks in the mind was somebody that was describing their feelings during COVID

00:15:44and the way that, certainly, again, in the UK traditional media, there would be a lot of emphasis on the way that hospitals were full.

00:15:53You know, they were full of patients, they were overflowing with patients and so on.

00:15:57And this particular participant was saying, well, at my local hospital, that's not the case.

00:16:03And so, what they were seeking from the news broadcasters was some indication that while it might indeed be the vast majority of hospitals that are full,

00:16:15nonetheless, an indication that it might not be all of them.

00:16:19So, I'm going to turn now to our second project.

00:16:28And we were keen to hear from people who, as I say, had previously held minority beliefs and had turned away from them.

00:16:37So, people tended to feel that others were more susceptible to sort of missing disinformation than they were.

00:16:46May I have just to say that because of this time, you have one minute left.

00:16:51Beg your pardon?

00:16:52In terms of time, you have one minute left.

00:16:54I have one minute left.

00:16:55Yes, exactly.

00:16:56That is pretty amazing.

00:16:58Right.

00:16:59Okay.

00:17:01I am going to flip through to the end.

00:17:03What a shame.

00:17:06Final slide.

00:17:08Core principles for messaging.

00:17:10This came out from our workshops.

00:17:11Firstly, the real vital importance of being aware of different circumstances, different people, different times.

00:17:21It's obvious, but it's really necessary.

00:17:23That it's, you know, the top-down approach, you know, might work at a kind of, you know, an initial level.

00:17:29But nonetheless, you really do need to delve into people's particular circumstances and what resonates for them.

00:17:37So, trusted voices are really important, ranging from huge kind of superstars, the voices of the national treasures of the UK, like David Attenborough and various others, all the way through to local community voices.

00:17:54Pharmacists in the area of health misinformation are really useful because they're so embedded and trusted within the community.

00:18:02And I think the final thing is absolutely this business about being non-judgmental and non-shaming.

00:18:12And so, the huge importance across all the people that we spoke to just about talking about it and talking about it in a kind of non-threatening, non-shaming way.

00:18:24And I think accepting, you know, hard though it is, that kind of saying, no, you're wrong, is really unlikely to move the dial.

00:18:34I'm so sorry. I ran out of time.

00:18:37That's perfect. Thank you very much.

00:18:43Thank you, Alison, for your talk on how people form and move away from certain beliefs.

00:18:47We turn now to the political side of the story, and our next speaker is Nicolas Jacquemais.

00:18:56Nicolas will explore how misinformation interacts with voting behavior and political attitudes.

00:19:02This is based on an experimental study that he lit during the last French presidential election.

00:19:08And where you look, Nicolas, actually, about how exposure to fake news can shape the way citizens think and make political choice.

00:19:19Nicolas, you're a professor at Paris Club Economics at the University of Paris 1.

00:19:23You're an honorary member of the Institut Universitaire de France,

00:19:25and you're also the head of the master's program at Paris Club Economics in Economics and Psychology.

00:19:32The floor is yours. Thank you.

00:19:35Thank you very much.

00:19:36I'm very excited to participate to this discussion.

00:19:40This is joint work with Laurence Vardazoglou, who, when we started this project, was a PhD student at PSC,

00:19:47and is now a brilliant consultant.

00:19:50And as the title says, the main focus of this paper is obviously the exposure to fake news and misinformation,

00:19:58and more specifically, what are the consequences of such an exposure.

00:20:02I will go fast on the background that I'm sure you all know even much better than I do,

00:20:08which is essentially that we do have some evidence that there are some consequences of being exposed to fake news.

00:20:16And so there are good reasons to think and believe in what I think we all have in mind,

00:20:20which is that the exposure to fake news and fake news circulation that we have seen happening in recent years

00:20:25is a key driver of the recent trends that we testify in our democracies.

00:20:31When we started to work on this project, it's been quite a big surprise to us to realize that, in fact,

00:20:38the one big question that we need to answer in order to get there,

00:20:41which is whether or not the exposure to fake news have consequences on the way people vote

00:20:47and the political choices that they make.

00:20:50This very question has received very little answer from the academic literature,

00:20:55and the academic evidence is scarce and typically not so convincing and, if anything, mixed.

00:21:02So once again, going a bit fast on the background,

00:21:05but there are a few things that are important to keep in mind for the reminder.

00:21:07First of all, observational evidence, based on what actually happened in terms of election

00:21:15after exposure to fake news, is not so convincing because typically exposure to fake news

00:21:20is a choice on the part of the individual, and we obviously expect particular people

00:21:24to choose to be exposed to fake news, and those people typically vote in a particular way,

00:21:29and so measuring the actual consequences of this exposure is quite difficult on observational data.

00:21:34There is still some evidence, and even this observational evidence is rather mixed.

00:21:39From a more convincing point of view, and in terms of experimental evidence,

00:21:44when I talk about experimental evidence here, I really refer to exposure to fake news

00:21:50that is decided independently from the characteristic and preferences of the individual,

00:21:57so as to measure the causal consequences on voting behavior of this exposure to fake news,

00:22:02that are essentially, to the best of my knowledge, of our knowledge, sorry to include you, Lorenz,

00:22:07and if you have any idea in mind, please share it with us.

00:22:11We found only two examples of a clear experimental evidence of the voting consequences of exposure to fake news.

00:22:18One is the U.S. with a small effect.

00:22:20I will focus more, obviously, on the one that has been run in France,

00:22:23with a few important aspects of the study to keep in mind.

00:22:28So this is a study that looks at the causal consequence of the exposure to fake news

00:22:34on the voting intentions during the presidential election that took place before the one we studied.

00:22:41And they look solely at voting intentions in favor of the extreme right party

00:22:48and do not elicit the full voting intentions for all possible parties that run for the election.

00:22:54And very importantly, they are able to disentangle the effect of fake news

00:22:59from what they call a saliency effect by having also treatments in which there is fact-checking,

00:23:04so people are not only exposed to the fake news but also to the fact-check of the fake news

00:23:09and some accurate news that all relate to the same topic, which is typically related to immigration.

00:23:15And the big result that they see is that they do find a positive effect on voting intentions in favor of the extreme right,

00:23:22but of all these three treatments, I know, and so it's not fake news on its own that, according to them,

00:23:27has an effect, but really the saliency of those political platforms that refer to particular parties,

00:23:34and so that's really the starting point of our study.

00:23:38As a matter of fact, the aim is to measure the effect of the exposure to fake news on voting intentions.

00:23:44For that, we used the last presidential election in France.

00:23:47We ran the study on a representative panel of the French population.

00:23:51And because we thought that this effect was already known,

00:23:54the very aim of the project was to go beyond the simple question of whether or not there are consequences

00:23:59and look in particular at whether the political attitudes that lie behind voting intentions

00:24:04are also affected by the exposure to fake news

00:24:07and whether the change in voting intentions is channeled by changes in political attitudes.

00:24:13You will see this is a big failure

00:24:14because essentially we don't see anything or very small effect in terms of voting intentions,

00:24:21which was once again a big surprise to us,

00:24:23which also means that essentially we, based on that, expect no much effect on voting on political attitudes,

00:24:32which indeed is the case.

00:24:34So just a few words on the important features of our study.

00:24:39So we were in the study shortly before the first round of the presidential election.

00:24:43The lineup of candidates were ordered according to groups to which I will refer later in the results section is on the slide.

00:24:53And the study was run on a representative sample of the adult population in age of voting

00:25:02in partnership with a big polling company in France.

00:25:06And we elicit in the study the voting intentions of these people

00:25:09in exactly the same way it is done by this polling company running so many different polls before the election takes place.

00:25:17The important feature of our study is that we will randomly, so based on a conflict,

00:25:24decide to which exposure our participants will participate.

00:25:30In the control there will be no exposure to particular news.

00:25:34And we will elicit, this is the left part, both the political attitudes,

00:25:38so what they think about different kinds of issues, what they think about the credibility of the candidates,

00:25:44what they think about the achievements of the incumbents.

00:25:48And then we will elicit their voting intentions for which candidate among the list they are more likely to vote.

00:25:55And we have some way to go beyond voting intentions

00:25:59and double-check that the voting intentions we elicit do say something about the way people actually intend to vote in the election.

00:26:06We can go back to that in the discussion if you want,

00:26:08but essentially there is a very large correlation between these different measures,

00:26:12and so voting intentions are really behaviorally grounded thanks to the design of the study.

00:26:17The two main treatments of interest are, first of all, a treatment in the middle,

00:26:21where we elicit political attitudes first,

00:26:25and conditional on these political attitudes,

00:26:27people, conditional, sorry,

00:26:29given these political attitudes,

00:26:31people move to the exposure of fake news,

00:26:33and conditional on this exposure to fake news,

00:26:35where they see political intentions,

00:26:37and so it gives us information about how exposure to fake news

00:26:42changes possibly voting intentions,

00:26:44conditional on what the political attitudes were before the exposure to fake news.

00:26:49This is the interest of the third treatments,

00:26:51in which the only change is that we expose participants to fake news

00:26:55before the elicitation of political attitudes.

00:26:58I'm very sorry.

00:27:00For me, the sound was not so good in the back, but okay.

00:27:04Which gives us information about the possible change in political attitudes

00:27:08that happens because of the exposure to fake news,

00:27:11and the resulting consequences on voting intentions.

00:27:14So essentially, we have two ways of analyzing the data that is produced

00:27:18by this experiment.

00:27:20Either we want to measure the effect of exposure to fake news

00:27:22on voting intentions,

00:27:24and we group the last two treatments,

00:27:26and compare them to the first,

00:27:28or we want to look at the effect of the exposure to fake news

00:27:31on political attitudes,

00:27:33and we group the first two,

00:27:34and compare them to the last.

00:27:39In terms of the exposure to fake news,

00:27:41and the main treatment we implemented,

00:27:43we used a set of fake news

00:27:45that actually circulated during this campaign.

00:27:48You can see two examples on these slides.

00:27:51They not only actually circulated,

00:27:52and are associated with clear sources,

00:27:56but they also have been fact-checked

00:27:58by sources I'm giving you

00:28:01on the bottom of the slide here.

00:28:03And so they have been identified as fake,

00:28:06and refer to typically populist narratives

00:28:08that again actually circulated during the campaign.

00:28:13And we ask a few questions to our subjects,

00:28:16to our participants, sorry,

00:28:17about this fake news,

00:28:19which I think are important to assess

00:28:23and ascertain the credibility of our results

00:28:27that I will show you first

00:28:29to give you a sense of the data.

00:28:31So first of all,

00:28:33we restrict the sample to participants

00:28:36who self-report.

00:28:38They are registered to vote,

00:28:40and so it's really impossible voters

00:28:42that will run the study.

00:28:43This is the table that you can see on the top.

00:28:45Essentially, there are no differences

00:28:46in this regard between treatments.

00:28:49And the bottom two figures are important for us

00:28:52because they give you,

00:28:53for each fake news,

00:28:55we have five of them in all treatments,

00:28:57in random order.

00:28:59For each fake news

00:29:00that takes a different color on the figure,

00:29:02I'm showing you on the left

00:29:04the answers of our participants

00:29:06about whether or not they think the claim,

00:29:09according to us, is true.

00:29:11And what this figure essentially says

00:29:13is that the choice of the fake news

00:29:15we expose our participants to

00:29:17is such that they do believe

00:29:19in the piece of information

00:29:21they are exposed to

00:29:22when we give it to them.

00:29:25And secondly, on the right-hand side,

00:29:27it is the summary of the answer

00:29:28to the question about whether or not

00:29:30they already saw this piece of information.

00:29:32If you run a fake news treatment

00:29:34and essentially expose people

00:29:35to some news they already had before,

00:29:37you don't measure the treatment effect

00:29:38because they already make decisions

00:29:41conditional on the exposure

00:29:43to this piece of information.

00:29:44So you can see that we have a large part

00:29:46of the sample

00:29:47who have not,

00:29:49according to what they self-report,

00:29:51seen the fake news before.

00:29:52And so it's really the exposure

00:29:54to this new piece of information

00:29:56that we measure in the experiment.

00:29:59And the timing is quite good

00:30:01because essentially I have two results

00:30:03to share with you.

00:30:05The first is in terms of voting intentions,

00:30:10what change in the candidates

00:30:13for whom our participants

00:30:14are more likely to vote

00:30:15according to their self-reported intentions,

00:30:18depending on whether they have been exposed

00:30:19to fake news.

00:30:21On the left,

00:30:22you can see the voting intentions

00:30:23grouped by the different sets

00:30:26of candidates I've shown you

00:30:27at the beginning.

00:30:28On the right, yes,

00:30:30you can see for each individual candidate,

00:30:33you can see that essentially

00:30:34the slight move in voting intentions

00:30:37that we can see happens

00:30:39in the very middle of the distribution,

00:30:42a slight increase in the voting intentions

00:30:44in favor of the center,

00:30:46which mainly benefits Macron.

00:30:49and this happens mainly

00:30:51at the expense of Mélenchon

00:30:53that you can see both

00:30:55in the extreme left

00:30:56on the left graph

00:30:58and as JLM on the right,

00:31:02for whom the loss is quite significant

00:31:05in terms of voting intentions

00:31:08and most of the increase

00:31:09in the center comes from the left

00:31:11and slightly also from the extreme right.

00:31:15So essentially in terms

00:31:16of voting intentions,

00:31:17what we see is that the exposure

00:31:19to populist narrative

00:31:20in the form of fake news

00:31:21slightly benefits center candidates.

00:31:25I will get back

00:31:26to a possible interpretation

00:31:27of that at the end

00:31:29of the presentation.

00:31:30This is not because of me,

00:31:31but it doesn't matter.

00:31:32So this is essentially

00:31:33to show you

00:31:34that what I said

00:31:35is statistically grounded.

00:31:37and next the question is

00:31:39whether or not

00:31:40this slight,

00:31:41very slight change

00:31:42in voting intention

00:31:43that we observe

00:31:44is grounded or not

00:31:46in political attitudes.

00:31:47We could also have changes

00:31:48in political attitudes

00:31:49that do not translate

00:31:50in terms of voting intentions.

00:31:53Once again,

00:31:54if anything,

00:31:54but the effects

00:31:55are typically small.

00:31:56I'm just giving you

00:31:57all the data

00:31:58to share the information

00:31:59and see how you react.

00:32:01On all issues

00:32:02that we elicit,

00:32:03we typically have

00:32:04very little change

00:32:06in political attitudes

00:32:07and if anything,

00:32:09changes in favor

00:32:10of more positive opinions

00:32:11of Macron

00:32:12and less agreement

00:32:14with the other candidates.

00:32:17So that's the comparison

00:32:18that you can see

00:32:19on the top of the slide.

00:32:21In terms of the achievements

00:32:22of the incubants,

00:32:23the story is typically

00:32:25very similar.

00:32:26There is a slight increase

00:32:27in how well people

00:32:29think Macron did

00:32:30once they have been exposed

00:32:31to this fake news

00:32:32I've shown you

00:32:33at the beginning.

00:32:34I'm done.

00:32:35The question is

00:32:36what can we do with that

00:32:38and what's the general story

00:32:39that these results say?

00:32:42First of all,

00:32:43the current existing evidence

00:32:45shows that it's essentially

00:32:47saliency that matters

00:32:48in terms of the consequences

00:32:49of the exposure

00:32:50to fake news.

00:32:52Saliency means

00:32:52you give more weight

00:32:53to particular issues.

00:32:56If saliency

00:32:56is really the main driver

00:32:58of the effect

00:32:58of the exposure

00:32:59to fake news

00:33:00on different kinds

00:33:00of behavior

00:33:01and voting intentions

00:33:02in particular,

00:33:02it could also well mean

00:33:05that the very threat

00:33:07of populist parties

00:33:09winning the election

00:33:10becomes more salient

00:33:12because of the exposure

00:33:13to fake news.

00:33:14That the saliency

00:33:15applies to, yes,

00:33:16the issues that are contained

00:33:17in the news

00:33:18but also to the threat

00:33:20of political voting,

00:33:22sorry,

00:33:22voting in favor

00:33:24of extreme parties.

00:33:25and so one

00:33:26of the possible interpretation

00:33:27is that strategic voting

00:33:29in favor

00:33:30of other candidates

00:33:32like in our case

00:33:33the extreme left

00:33:33could be,

00:33:35not strategic voting,

00:33:36sorry,

00:33:37sincere voting

00:33:37in favor of the extreme left

00:33:39could be disciplined

00:33:39by the exposure

00:33:40to fake news

00:33:41with more strategic voting

00:33:42that goes

00:33:43to the one candidate

00:33:44that is more likely

00:33:45to win

00:33:45given the risk

00:33:47that became more salient

00:33:48because of the exposure

00:33:49to fake news.

00:33:50The general story

00:33:51that I think

00:33:52we can draw

00:33:55from this result

00:33:56and again,

00:33:57the lack of convincing evidence

00:33:59of a huge

00:34:00and significant effect

00:34:01of the exposure

00:34:02to fake news

00:34:02and fake news circulation

00:34:03on voting intentions

00:34:05is also an invitation,

00:34:07I think,

00:34:08to reflect back

00:34:09on why we give

00:34:11so much weight

00:34:12to the circulation

00:34:13of fake news

00:34:14in this trend

00:34:15I was talking about

00:34:16at the beginning.

00:34:18The idea that fake news

00:34:19is the driver

00:34:20and the main reason

00:34:21why we observe

00:34:22what we observe

00:34:23relies on the idea

00:34:24and this has been

00:34:25illustrated by the talk

00:34:26just before

00:34:27that people are

00:34:28typically misinformed,

00:34:29typically unable

00:34:30to reach

00:34:31for the information

00:34:32that is more accurate

00:34:33and typically unable

00:34:35to disentangle

00:34:36fake news

00:34:37from accurate information.

00:34:40In a sense,

00:34:41it could well be

00:34:42that this is the tree

00:34:44that hides the forest

00:34:45that in fact

00:34:46fake news circulation

00:34:47is not on its own

00:34:49the one thing

00:34:49we need to fight

00:34:50against

00:34:51in order to counter

00:34:52the many

00:34:53very serious issues

00:34:55that we are facing

00:34:56but rather

00:34:58more deeply

00:34:58the way people

00:35:00seek for information

00:35:01and the political opinions

00:35:04and other drivers

00:35:06that in fact

00:35:07lie behind

00:35:07both the strengths

00:35:08that we observe

00:35:09and the reason why

00:35:12there is more

00:35:13and more fake news

00:35:14circulation.

00:35:15I think I'm not too bad

00:35:16in terms of timing

00:35:17and thank you very much.

00:35:25Perfect one time.

00:35:26Thank you, Nicolas.

00:35:27Yeah, it's very interesting

00:35:28and especially

00:35:29we are in the time

00:35:30with a lot of uncertainty

00:35:31in terms of politics

00:35:32and next national election

00:35:35is in one year and a half

00:35:36at most

00:35:37so your analysis

00:35:38is priceless

00:35:39and meaningful.

00:35:40we keep going

00:35:43with elections

00:35:44but we shift our focus

00:35:45to the platforms themselves

00:35:46and I turn now

00:35:47to Pedro Ramaciotti

00:35:49Pedro, you are

00:35:50CNRS Research Fellow

00:35:52at the Complex

00:35:53System Institute of Paris

00:35:54and Professor

00:35:55at the Media Lab

00:35:56in Sciences Po

00:35:57your work

00:35:59cover a wide range

00:36:00of areas

00:36:01and actually

00:36:02what you are going

00:36:03to present today

00:36:03is a perfect reflection

00:36:05of that.

00:36:07You will take us

00:36:08into

00:36:09fact-checking

00:36:11systems

00:36:12such as

00:36:13community notes

00:36:14on Twitter

00:36:15and

00:36:16you will explore

00:36:18how social media

00:36:19companies are trying

00:36:19to manage

00:36:20misinformation

00:36:21and its implications

00:36:23for election

00:36:23and democratic integrity

00:36:25so Pedro

00:36:25the floor is yours.

00:36:27Thank you.

00:36:29So I have this

00:36:30very long title

00:36:30which I try

00:36:34to summarize here

00:36:35in a sense

00:36:35that will become

00:36:36hopefully more clear

00:36:37which points to message

00:36:39crowd-checking

00:36:41is not enough

00:36:41and I'm going to try

00:36:43to be very schematic

00:36:44so first I'm going to

00:36:45tell a little bit

00:36:47how we've been moving

00:36:48more and more

00:36:48from expert fact-checking

00:36:50to crowd-sourced fact-checking

00:36:52and what it means

00:36:52and then to present

00:36:54a very prevalent system

00:36:56that are being deployed

00:36:58in X

00:36:59but also copied

00:37:00by other platforms

00:37:01and the inner workings

00:37:02of that

00:37:03what's going on

00:37:03inside the algorithms

00:37:04that govern

00:37:05these systems

00:37:06and then to present

00:37:07the court

00:37:08of the presentation

00:37:08which is

00:37:09a large-scale study

00:37:11that we conducted

00:37:11showing how the system works

00:37:14spoiler it works

00:37:15quite well

00:37:16but it leads

00:37:17to a problem

00:37:19an additional problem

00:37:20that is important

00:37:21for regulation

00:37:21so these crowd-checking

00:37:24systems work

00:37:26in a way

00:37:27in which

00:37:27you have a post

00:37:28and regular users

00:37:29can provide moderation

00:37:31to posts

00:37:31such as in here

00:37:32you have a note

00:37:33that is not written

00:37:34by a professional fact-checker

00:37:35but just a regular user

00:37:36and the reason

00:37:38why these things

00:37:39have been tested

00:37:40is because there are

00:37:41some limitations

00:37:41with expert fact-checking

00:37:43so there's a problem

00:37:44of volume

00:37:45there's an amount

00:37:46of posts

00:37:47that can be checked

00:37:48a reach

00:37:49into niches

00:37:51linguistic

00:37:52community niches

00:37:53thematic niches

00:37:54there's also a question

00:37:55of reactivity

00:37:56and speed

00:37:57so it needs to go

00:37:58through quite a

00:37:59sometimes long loop

00:38:01it would be flagged

00:38:02in a platform

00:38:02then sent to expert

00:38:03fact-checkers

00:38:04sometimes go into

00:38:05partner institutions

00:38:06that would check

00:38:06send back to platforms

00:38:07that would then activate

00:38:09or not some

00:38:10moderating action

00:38:11and then there are

00:38:13two related issues

00:38:13one is the degree

00:38:14of trust

00:38:15that would put

00:38:16in these institutions

00:38:17that conduct

00:38:18this expert fact-checking

00:38:20and also how

00:38:21this trust

00:38:22or lack thereof

00:38:23can further reach

00:38:25a state in which

00:38:27people distrusting

00:38:28this mechanism

00:38:29can further entrench

00:38:30in beliefs

00:38:32generated

00:38:34by misleading

00:38:34misinformation

00:38:36or just prompting

00:38:38them to share more

00:38:39this has been shown

00:38:40in research

00:38:40so there have been

00:38:42some alternatives

00:38:43and one of the

00:38:43alternatives is

00:38:44crowd-sourced

00:38:46fact-checking

00:38:46a pioneer

00:38:48in this space

00:38:48is Birdwatch

00:38:49deployed in the

00:38:50United States

00:38:50by Twitter

00:38:51in 2021

00:38:51then deployed

00:38:53worldwide

00:38:54with this different

00:38:54name

00:38:55community notes

00:38:56and it works

00:38:57like this

00:38:58there's a post

00:38:59and then

00:38:59someone can attach

00:39:01a note

00:39:01and then

00:39:02crucially

00:39:02people vote

00:39:03on the moderation

00:39:04action

00:39:04saying

00:39:05this moderation

00:39:07action

00:39:08was useful

00:39:08not useful

00:39:09and then

00:39:10the platform

00:39:11will need to

00:39:12gather all this

00:39:13voting

00:39:13and decide

00:39:14whether

00:39:15different notes

00:39:17are well

00:39:17suited or not

00:39:19to be shown

00:39:19and it will pick

00:39:21one

00:39:21and put it

00:39:22to be shown

00:39:22to everyone

00:39:23visiting

00:39:23or reading

00:39:24that post

00:39:24so what they

00:39:27claim

00:39:27they publish

00:39:29a paper

00:39:29and this

00:39:29is that this

00:39:31reduces

00:39:31by a fourth

00:39:33the probability

00:39:34that someone

00:39:35will believe

00:39:36on the misleading

00:39:36claim

00:39:37also by a fourth

00:39:39or a third

00:39:39the probability

00:39:40that someone

00:39:40will share

00:39:41the misleading

00:39:41claim

00:39:42other researchers

00:39:43have found

00:39:44that this is

00:39:44quite fast

00:39:45so half

00:39:46of the posts

00:39:48that have

00:39:48these notes

00:39:48reach

00:39:50some

00:39:51moderation

00:39:52action

00:39:53in 15 hours

00:39:55or three

00:39:56or four hours

00:39:57from the

00:39:57moderation

00:39:57action

00:39:58also

00:39:59speaking to

00:40:00the keynote

00:40:01before

00:40:01platforms

00:40:02have been

00:40:03feeling

00:40:04more authorized

00:40:05to advance

00:40:06towards the system

00:40:06because it's

00:40:07aligned

00:40:07better with

00:40:09the new

00:40:09US administration

00:40:10on what

00:40:11freedom of speech

00:40:12is and

00:40:12governance

00:40:13of platform

00:40:14is

00:40:14and of course

00:40:15this has

00:40:15reduced

00:40:16costs

00:40:16so

00:40:17now I want

00:40:18to talk

00:40:18about the

00:40:19question of

00:40:19how does

00:40:20this system

00:40:21work

00:40:21specifically

00:40:21the algorithms

00:40:22that select

00:40:23the moderation

00:40:24node that is

00:40:25shown to users

00:40:26so

00:40:27this is the

00:40:28problem

00:40:29that the

00:40:29engineers

00:40:30and researchers

00:40:30in the

00:40:31platforms

00:40:31were trying

00:40:31to solve

00:40:32and

00:40:33yeah

00:40:35so

00:40:35naive approaches

00:40:36will likely

00:40:37fail

00:40:38because it's

00:40:39been shown

00:40:39that

00:40:39polarization

00:40:40is primarily

00:40:41driven

00:40:41by

00:40:42sorry

00:40:43the sharing

00:40:44of misinformation

00:40:45is

00:40:45primarily

00:40:46driven

00:40:46by polarization

00:40:47so if you

00:40:48don't take

00:40:49this into

00:40:49account

00:40:49you can

00:40:50fall into

00:40:51a situation

00:40:51in which

00:40:52someone

00:40:52across the

00:40:53political

00:40:53spectrum

00:40:54will produce

00:40:55a misleading

00:40:55claim

00:40:56that will

00:40:57get a

00:40:57moderation

00:40:57note

00:40:58but people

00:40:59from across

00:41:00the spectrum

00:41:01will find it

00:41:02useful as a

00:41:03moderation

00:41:03action

00:41:03but people

00:41:04from the

00:41:05side side

00:41:05of the

00:41:06moderation

00:41:06the ideological

00:41:08spectrum

00:41:08will find it

00:41:09not as useful

00:41:09so you can

00:41:10fall into a

00:41:10situation

00:41:11in which

00:41:11you have

00:41:11if you

00:41:12will

00:41:13echo

00:41:13chambered

00:41:14moderation

00:41:15actions

00:41:15so what

00:41:17they propose

00:41:18is to

00:41:19decompose

00:41:20the weight

00:41:22that the

00:41:22ideology

00:41:23might have

00:41:23in how

00:41:24people

00:41:25vote on

00:41:25these

00:41:26notes

00:41:26that are

00:41:26proposed

00:41:27so they're

00:41:28looking at

00:41:28data that

00:41:29looks like

00:41:29this

00:41:29so they have

00:41:30the notes

00:41:31that have

00:41:31been proposed

00:41:31and the

00:41:32people have

00:41:33been voting

00:41:33on them

00:41:34and they

00:41:34attribute

00:41:35numerical

00:41:35values to

00:41:36that

00:41:36one being

00:41:37voted

00:41:38useful

00:41:38zero

00:41:38not useful

00:41:39and things

00:41:40in between

00:41:40and here

00:41:41I'm going

00:41:41to show

00:41:41an equation

00:41:42but I think

00:41:42it's going

00:41:42to be useful

00:41:43and very

00:41:43simple

00:41:44so what

00:41:45they try

00:41:45to do

00:41:45is to

00:41:46look at

00:41:46all of

00:41:47this data

00:41:47how people

00:41:48are voting

00:41:48on the

00:41:49usefulness

00:41:50on these

00:41:50moderation

00:41:51actions

00:41:51and to

00:41:52model this

00:41:53in a way

00:41:53that decomposes

00:41:54different things

00:41:55that are

00:41:55important

00:41:55the thing

00:41:56the first

00:41:57thing is

00:41:57a platform

00:41:59baseline

00:42:00the second

00:42:01one is

00:42:01the contribution

00:42:02that comes

00:42:02only from

00:42:03the people

00:42:03that vote

00:42:04the note

00:42:06so maybe

00:42:06people are

00:42:06more lenient

00:42:07different people

00:42:07are more lenient

00:42:08and vote

00:42:09more positively

00:42:09on notes

00:42:10then a part

00:42:11that is only

00:42:11attributable

00:42:12to the moderation

00:42:13action

00:42:13to the note

00:42:14and then

00:42:15and this is

00:42:16the crucial part

00:42:16a part

00:42:17that depends

00:42:18on the ideological

00:42:19alignment

00:42:20between the people

00:42:21that produced

00:42:21the moderation

00:42:22action

00:42:22the note

00:42:23and the people

00:42:24that voted

00:42:25that note

00:42:25so they

00:42:26hope to capture

00:42:27the problematic

00:42:28part in there

00:42:29and this is

00:42:31what it leads

00:42:31like when we

00:42:32run this data

00:42:33and fit the model

00:42:34and look at

00:42:35these parameters

00:42:35these are

00:42:37the notes

00:42:38these clouds

00:42:39are density

00:42:40of notes

00:42:41shown along

00:42:41two of these

00:42:42parameters

00:42:43and the horizontal

00:42:44axis you have

00:42:45the ideology

00:42:45so going from

00:42:46left to right

00:42:47and on the

00:42:48vertical axis

00:42:49you have this

00:42:49intrinsic usefulness

00:42:51of the note

00:42:51according to the

00:42:52people that voted

00:42:53and here in these

00:42:54three groups

00:42:55it's what the

00:42:56platform has done

00:42:57with these notes

00:42:58so on the left

00:43:00you have what

00:43:01platforms has taken

00:43:02with this algorithm

00:43:03and said

00:43:03this is useful

00:43:04this is the ones

00:43:05that we are going

00:43:06to select

00:43:07and show to

00:43:07everyone

00:43:08next to the post

00:43:09right

00:43:09so what they're

00:43:10doing in essence

00:43:11is putting a

00:43:12threshold on this

00:43:13parameter of

00:43:14intrinsic usefulness

00:43:15that is independent

00:43:16of the ideology

00:43:17the ideological

00:43:18alignment between

00:43:19the note

00:43:20and the person

00:43:20that rated the note

00:43:21then there are

00:43:22other status

00:43:23for example

00:43:23when these parameters

00:43:25threshold is not

00:43:26rich

00:43:26it means that it's

00:43:27quite divisive

00:43:28so there's no

00:43:28consensus

00:43:29and there's sometimes

00:43:30consensus on that

00:43:31the note is

00:43:32awful

00:43:33right

00:43:33so it looks

00:43:35concretely something

00:43:37like this

00:43:37when a note

00:43:38gathers consensus

00:43:39across the spectrum

00:43:40it will be

00:43:41evaluated positively

00:43:42and the platform

00:43:43cuts on this level

00:43:44of intrinsic usefulness

00:43:45across the political

00:43:46spectrum

00:43:47and the others

00:43:48look like this

00:43:49what I referred

00:43:50before as

00:43:51echo chamber

00:43:52kind of effect

00:43:52of moderation

00:43:53and consensus

00:43:54on the low quality

00:43:56of the note

00:43:56so does this

00:43:59system work

00:44:00this is the main

00:44:01topic of a study

00:44:02we conducted

00:44:03with Paul Bouchot

00:44:04in which we look

00:44:05at how the system

00:44:06performs outside

00:44:07the United States

00:44:08across the world

00:44:09because most results

00:44:11are evaluated

00:44:12in the United States

00:44:13also because

00:44:14this critical

00:44:15parameter

00:44:16about ideology

00:44:17it's designed

00:44:18with the US

00:44:19ideological

00:44:20cleavages in mind

00:44:21so the question

00:44:23is how does it work

00:44:23around the world

00:44:24and what fails

00:44:25to get moderated

00:44:26this way

00:44:26so all of this

00:44:29data that we're

00:44:29using is open

00:44:30they open this data

00:44:32anonymized

00:44:32except that

00:44:33the contextual

00:44:34data is not

00:44:35if you want to

00:44:35link a note

00:44:36to a post

00:44:37and the post

00:44:37to someone

00:44:38this requires

00:44:39you have API access

00:44:40which is quite

00:44:41expensive

00:44:42or article 40

00:44:43access

00:44:44so this is

00:44:45all the data

00:44:46that existed

00:44:47in the system

00:44:48up until the date

00:44:49we took for analysis

00:44:51which is March

00:44:52of this year

00:44:52so we have over

00:44:53two years

00:44:53and the first thing

00:44:54to notice

00:44:55is that

00:44:56in all

00:44:57there have been

00:44:57130 million ratings

00:45:00given to

00:45:012 million notes

00:45:01produced by

00:45:02more than

00:45:03a million users

00:45:03which is not

00:45:05huge in a sense

00:45:07right

00:45:07the other thing

00:45:09that to notice

00:45:10is that

00:45:10as in any

00:45:11online system

00:45:12you have a lot

00:45:13of concentration

00:45:14so 1% of note

00:45:16orders

00:45:17have written

00:45:17a third of the notes

00:45:19a third of the notes

00:45:20have been given

00:45:20to only a thousand

00:45:21of accounts

00:45:22that gather

00:45:23most of this

00:45:24moderation

00:45:25attention

00:45:25by the public

00:45:26one account

00:45:28has been requesting

00:45:29notes

00:45:31for post

00:45:32most of the

00:45:33for most of the post

00:45:34and there's high

00:45:35concentration

00:45:36between countries

00:45:37so most of the notes

00:45:39are in the United States

00:45:40but actually

00:45:40when you look

00:45:41at the proportion

00:45:42compared

00:45:43with the number

00:45:44of users

00:45:44in the platform

00:45:45it reveals

00:45:46wide adoption

00:45:48in Europe

00:45:49or at least

00:45:50homogeneous adoption

00:45:51in Europe

00:45:51also users

00:45:54that provide

00:45:55these notes

00:45:55are using

00:45:56sources

00:45:57as expert

00:45:58fact checkers

00:45:59do

00:45:59most of the time

00:46:00is news media articles

00:46:01sometimes it would be

00:46:02post

00:46:03other posts

00:46:04from the platform

00:46:05sometimes Wikipedia

00:46:06articles

00:46:07and to a very lesser extent

00:46:08fact checking articles

00:46:11so the things

00:46:13that are discussed

00:46:14in this post

00:46:15that people are moderating

00:46:16is mostly

00:46:17politics

00:46:18and this we've measured

00:46:19with different

00:46:20topic modeling techniques

00:46:21also this attracts

00:46:23most of

00:46:24of the voting action

00:46:26that people take

00:46:26towards this note

00:46:27and this political content

00:46:31is spread

00:46:32across different

00:46:33sub themes

00:46:34people to discuss

00:46:35a lot about

00:46:36candidates

00:46:37parties

00:46:38authorities

00:46:38but also about

00:46:39crime, police

00:46:40and the judiciary

00:46:41redistribution

00:46:43taxes

00:46:44poverty

00:46:44the environment

00:46:45and freedom

00:46:46and human rights

00:46:46so the main

00:46:48questions here

00:46:49is

00:46:49if

00:46:50what interests

00:46:51the people

00:46:52that are doing

00:46:52the moderation

00:46:53is

00:46:53politics

00:46:55and a key

00:46:56component

00:46:56in the algorithm

00:46:57algorithmic decision

00:46:58on deciding

00:46:59what to show

00:47:00is ideological

00:47:01what is it

00:47:02that is measured

00:47:04as ideology

00:47:04around the world

00:47:06outside the United States

00:47:07so to do this

00:47:08what we did

00:47:08is

00:47:09we took

00:47:10the users

00:47:11involved in this

00:47:12exercise

00:47:13and inferred

00:47:14for them

00:47:14ideological positions

00:47:15on different scales

00:47:16so we started

00:47:19with the United States

00:47:20as a baseline

00:47:20so we know

00:47:21the ideological position

00:47:23of the person

00:47:24that received

00:47:25the moderation action

00:47:26and the ideological

00:47:26position perceived

00:47:28by this community

00:47:28notes algorithm

00:47:29in moderating

00:47:31and it works

00:47:32quite well

00:47:33in the sense

00:47:33that notes

00:47:34that are perceived

00:47:35as left-leaning

00:47:36by the algorithm

00:47:37are given

00:47:37to right-leaning people

00:47:39and the mirror effects

00:47:41exist on the left

00:47:42and on the rater side

00:47:44so people

00:47:44that are rating

00:47:45these notes

00:47:45are also very coherent

00:47:47ideologically

00:47:47so people that

00:47:48according to the algorithm

00:47:49are left-leaning

00:47:50are rating positively

00:47:52notes given

00:47:52to right-leaning people

00:47:54and the mirror effects

00:47:55exist also

00:47:57on the other side

00:47:57of the spectrum

00:47:58but the question is

00:47:59how does it work

00:48:00outside the United States

00:48:02where the question

00:48:02of ideology

00:48:04it's quite different

00:48:05might be organized

00:48:06across different dimensions

00:48:08so we measure this

00:48:08across several dimensions

00:48:09and here we present

00:48:10results for two dimensions

00:48:12left-right

00:48:14and a dimension

00:48:15that is important

00:48:16also misinformation

00:48:16which is attitudes

00:48:18towards elites

00:48:19and institutions

00:48:20so in here

00:48:20we have the same systems

00:48:22the same systems

00:48:24in these two dimensions

00:48:25and critically

00:48:26what we measure here

00:48:27is this delta-1 arrows

00:48:29are what we know

00:48:31is the main structuring

00:48:32combination of these

00:48:33two dimensions

00:48:33in these different countries

00:48:34and CN

00:48:36the community notes

00:48:36is what the algorithm

00:48:37thinks is measuring

00:48:38and it's quite aligned

00:48:39for example in France

00:48:40it's some combination

00:48:42of the two

00:48:42so we look at it

00:48:44in different countries

00:48:45and it works

00:48:47in capturing

00:48:47the important dimensions

00:48:49across the world

00:48:50even though it was not

00:48:50explicitly conceived

00:48:52for that

00:48:52and it works

00:48:53in the sense

00:48:54that people across

00:48:55the spectrum

00:48:55are providing notes

00:48:56and rating them

00:48:57there's a big debate

00:48:59about asymmetries here

00:49:01to which we provide

00:49:03a lot of caveats

00:49:03and counterfactuals

00:49:04so the fact

00:49:06that it's working

00:49:06so well

00:49:07it's also a problem

00:49:08because for all

00:49:10the disadvantages

00:49:11that the system

00:49:12might have

00:49:12so it's opacity

00:49:13in auditing

00:49:14the cost of auditing

00:49:16it also excludes

00:49:17some national actors

00:49:18from the moderating exercise

00:49:19of course

00:49:20has a lot of benefits

00:49:21in cost

00:49:22for the platform

00:49:24providing the moderation

00:49:25but it comes

00:49:27with one inner limitation

00:49:29that our study

00:49:30highlights

00:49:30which is that

00:49:31by design

00:49:32this algorithm

00:49:34called bridging algorithms

00:49:35are made to surface

00:49:36what generates consensus

00:49:37but factual information

00:49:40that is polarizing enough

00:49:42cannot generate consensus

00:49:44and this is of course

00:49:45a problem to content

00:49:46that activates

00:49:47this polarization

00:49:48and in the context

00:49:50of election

00:49:50this is a problem

00:49:51and election

00:49:51is one of the key things

00:49:53that platforms

00:49:53should look for

00:49:54in quality of content

00:49:56when they produce

00:49:56self-assessments

00:49:58so to show this

00:49:59we have measured

00:50:00the degree

00:50:01to which

00:50:02these notes

00:50:03given to different topics

00:50:05compared to

00:50:06topics of elections

00:50:08in the same period

00:50:09for four key elections

00:50:10that occurred

00:50:11during our time

00:50:12of observation

00:50:12which is

00:50:13the presidential

00:50:14in the United States

00:50:15the general election

00:50:15in the United Kingdom

00:50:17the legislative in France

00:50:19and the federal election

00:50:20in Germany

00:50:21and it's consistently

00:50:22under a moderate

00:50:23so the key thing here

00:50:26is that

00:50:27moving forward

00:50:29in the regulatory discussion

00:50:32it would be key

00:50:33to distinguish the systems

00:50:34not all of these systems

00:50:35are equal

00:50:36in what crowdsources means

00:50:38what are the regular users

00:50:39that are allowed to do this

00:50:40when and what sequence

00:50:42and what are the algorithms

00:50:43that are driving this behind

00:50:45the second thing is that

00:50:47this of course

00:50:48is not all negative

00:50:49it has quite a lot of advantages

00:50:51speed and reach

00:50:53is one of those

00:50:54these new architectures

00:50:56will move

00:50:57will change

00:50:58the delimitation

00:50:59of the agency

00:51:00of different actors

00:51:01especially national actors

00:51:02in participating

00:51:03in the moderation loop

00:51:04and the main thing

00:51:06that we want to stress

00:51:08sometimes

00:51:09with the study

00:51:10is that

00:51:10these are not neutral

00:51:11so for whatever

00:51:12discourse

00:51:13that is put forward

00:51:14on people

00:51:15doing the regulation

00:51:16this comes

00:51:17with a lot of knobs

00:51:19that the platform

00:51:20needs to adjust

00:51:21in parameters

00:51:21and design

00:51:22that completely change

00:51:23the outcome

00:51:24of the moderation

00:51:25the other key thing

00:51:27that we highlight

00:51:29in our study

00:51:30is that it comes

00:51:30with a fundamental limitation

00:51:32so these kind of algorithms

00:51:34cannot address

00:51:35the most polarizing content

00:51:37it will lead to content

00:51:40that will go

00:51:40under moderated

00:51:42so it would point

00:51:43to a state of moderation

00:51:46in which maybe

00:51:46a combination of both

00:51:48might be beneficial

00:51:50in getting the best

00:51:50of both worlds

00:51:51thank you

00:51:52and this is the

00:51:52reference for the study

00:51:53if you want to

00:51:54go take a look

00:51:55thank you Pedro

00:52:03to conclude this session

00:52:05in our morning

00:52:06so we just heard

00:52:07about

00:52:07how to design

00:52:09fast checking system

00:52:10and algorithm

00:52:11and we keep going

00:52:12with a new generation

00:52:13of fast checking tools

00:52:15powered by

00:52:15artificial intelligence

00:52:16our last speaker

00:52:18is Thomas Renaud

00:52:19Thomas is professor

00:52:20of economics

00:52:20at the Université Paris-Saint-Claire

00:52:21and his research

00:52:22focused on the role

00:52:23of information

00:52:24and communication

00:52:24in shaping opinions

00:52:26and economic decisions

00:52:27so we are going

00:52:30to dive now

00:52:31in something

00:52:32actually I didn't know

00:52:33so that the word

00:52:34of the large language

00:52:35model boat

00:52:36used for fast checking

00:52:37on social media

00:52:38and Thomas

00:52:40you will discuss

00:52:41how people

00:52:42are turning

00:52:43to rely

00:52:43on this technology

00:52:45to verify information

00:52:46in real time

00:52:46and especially

00:52:47around political

00:52:48and breaking news events

00:52:49so I think

00:52:50it's a perfect

00:52:51compliment

00:52:52with what we just hear

00:52:53thank you

00:52:55thank you

00:52:56so that's a joint

00:52:57work with

00:52:58Mohsen Mosley

00:53:00from Oxford University

00:53:01and David Ren

00:53:02that now moves

00:53:03to Cornell

00:53:04and it's the first

00:53:06public presentation

00:53:07of those results

00:53:08so of course

00:53:09any comments

00:53:10are welcome

00:53:11so if you

00:53:13are on X

00:53:14and if you haven't

00:53:15quit X

00:53:16in recent year

00:53:18you may have seen

00:53:19that now users

00:53:20anyone

00:53:21can now tag

00:53:23large language

00:53:24model boats

00:53:25directly within

00:53:26the platforms

00:53:27to ask

00:53:28any question

00:53:29so if you have

00:53:30WhatsApp

00:53:30you may also have seen

00:53:31that there's a

00:53:32WhatsApp boat

00:53:32on which you can

00:53:33ask some stuff

00:53:33directly in WhatsApp

00:53:34there's now the same

00:53:36thing since March

00:53:37this year

00:53:38on X

00:53:39and there's two

00:53:41main boats

00:53:42that are now

00:53:43available

00:53:44which are Grok

00:53:45which is the one

00:53:46designed by

00:53:47XAI

00:53:49so from Elon Musk

00:53:50and there's another

00:53:51one from a company

00:53:52called Perplexity

00:53:54so in large language

00:53:55model of course

00:53:55you have ChatGPT

00:53:56you have Gemini

00:53:57you have Claude

00:53:58you have Grok

00:53:58and we have Perplexity

00:54:00and we are going to

00:54:01work on both

00:54:02Grok and Perplexity

00:54:03because the two

00:54:04have created a boat

00:54:06on X

00:54:07that allows any user

00:54:09to tag them

00:54:10in any tweets

00:54:11or any replies

00:54:12and ask them anything

00:54:13and the boat

00:54:14will answer

00:54:15directly within

00:54:16the platform

00:54:17and we're going

00:54:18to study that

00:54:19because as I'm

00:54:20going to show you

00:54:21it's used also

00:54:22widely for

00:54:23white checking

00:54:24for fact checking

00:54:26sorry

00:54:26within the platform

00:54:27so the goal

00:54:28of this topic

00:54:29and to the best

00:54:30of our knowledge

00:54:30it's really the first

00:54:31paper that's going

00:54:32to document

00:54:32the scale

00:54:33the topics

00:54:34and partisan patterns

00:54:36of LLM

00:54:37fact checking

00:54:38requests

00:54:38within the platform

00:54:39so we're going

00:54:41to quantify

00:54:41the agreement

00:54:42between

00:54:43different LLMs

00:54:44and evaluate

00:54:46inter LLM agreement

00:54:47we're going to

00:54:48also compare LLMs

00:54:49with community

00:54:50notes

00:54:51and discuss

00:54:52if there's any

00:54:52like complementarities

00:54:53or trade-off

00:54:54and what's very

00:54:55important

00:54:56we're going to

00:54:57try to see

00:54:57if the answer

00:54:59from the LLMs

00:55:00are correct

00:55:00or not

00:55:01and we'll compare

00:55:02that with

00:55:03professional fact

00:55:04checker ratings

00:55:05with two

00:55:06now three

00:55:06fact checkers

00:55:07that we basically

00:55:08pay for

00:55:08to classify

00:55:09a random subset

00:55:10of claims

00:55:11on which we have

00:55:12answers from

00:55:13an LLM

00:55:13so that's typically

00:55:15the type of

00:55:16interaction

00:55:17that we're going

00:55:18to analyze

00:55:18so on the left

00:55:21you have an example

00:55:21of a tweet

00:55:22which is a tweet

00:55:23saying that

00:55:24researchers found

00:55:25that COVID

00:55:25can destroy

00:55:26female eggs

00:55:28and after this tweet

00:55:29which was widely

00:55:30viewed

00:55:31you can see

00:55:312.6 million views

00:55:32there's someone

00:55:33that simply

00:55:35replied to the tweet

00:55:36mentioning growth

00:55:37and asking

00:55:38is this true

00:55:40after

00:55:41and it's often

00:55:42within

00:55:43one or two minutes

00:55:44there's an automatic

00:55:45answer from growth

00:55:47and in this case

00:55:48the answer says

00:55:49that no

00:55:49there's no evidence

00:55:50that supports

00:55:51the claim

00:55:51that COVID-19

00:55:52vaccines

00:55:53damage

00:55:54female fertility

00:55:56on the right

00:55:57you have another

00:55:58example

00:55:58related to

00:56:00economics

00:56:00and China's

00:56:01tariff

00:56:02on which also

00:56:03someone

00:56:03asked for

00:56:05a fact checking

00:56:06from growth

00:56:06and we have

00:56:07this answer

00:56:08in that case

00:56:08the LLM

00:56:10says that

00:56:10yes it's correct

00:56:12so what you can see

00:56:14and that's going to be

00:56:15important for what

00:56:16we're going to study

00:56:16is that

00:56:17this is really a reply

00:56:18so it's a public reply

00:56:20so we can observe

00:56:21everything on the sequence

00:56:23including the content

00:56:25of the initial tweet

00:56:26the political leaning

00:56:28of the initial posters

00:56:29and the ones

00:56:30that request a note

00:56:31plus some information

00:56:32that will derive

00:56:33from the answer

00:56:34from the LLM

00:56:35but you can see

00:56:36that if you compare

00:56:37the number of views

00:56:38of the initial claim

00:56:39compare the number

00:56:39of views

00:56:40of the fact check

00:56:41request

00:56:42or the answer

00:56:42from grok

00:56:43number of views

00:56:45is really small

00:56:46and that's

00:56:47why we can also

00:56:49see a difference

00:56:50with community notes

00:56:50because it really

00:56:51stems from

00:56:52a genuine demand

00:56:54for information

00:56:54rather some

00:56:55strategic motive

00:56:57that could have

00:56:58been documented

00:56:59for community notes

00:57:00so what do we do

00:57:02in this paper

00:57:03so we have collected

00:57:04all tweets

00:57:05mentioning grok

00:57:07or at perplexity

00:57:08up to july 6

00:57:102025

00:57:11so it was just

00:57:13before the launch

00:57:14of grok 4

00:57:15so the new model

00:57:16from X

00:57:18we are now

00:57:18currently extending

00:57:19the time period

00:57:20to also include

00:57:21grok 4

00:57:22and see for any

00:57:22differences

00:57:23between grok 3

00:57:24and grok 4 model

00:57:26and from those

00:57:2732 million tweets

00:57:29mentioning an LLM

00:57:31we identify

00:57:32specifically fact

00:57:33checking requests

00:57:34so those are tweets

00:57:35that are replies

00:57:37to an initial claim

00:57:38and that mention

00:57:39an LLM

00:57:39containing some

00:57:40patterns such as

00:57:41is it true

00:57:42is this real

00:57:43please fact check

00:57:44true or false

00:57:45verify this

00:57:46etc

00:57:46and at the end

00:57:48we end up with

00:57:49more than 500,000

00:57:51observations

00:57:51and one observation

00:57:53is a three tweet

00:57:54sequence

00:57:54an initial claim

00:57:56from anyone

00:57:57someone asking

00:57:58a fact check

00:57:59request

00:57:59and the answer

00:58:01from the LLM

00:58:02either from grok

00:58:03or for perplexity

00:58:05so about the scale

00:58:07and adoption

00:58:08so is it something

00:58:10minor

00:58:10and why do we think

00:58:11it's really worth

00:58:12investigating

00:58:13these nutrients

00:58:14first when we look

00:58:15at all tweets

00:58:16mentioning grok

00:58:17we find that

00:58:18more than 5%

00:58:20of those tweets

00:58:20are direct

00:58:22fact check requests

00:58:23other usage

00:58:25could include

00:58:26like asking

00:58:27for more explanation

00:58:28asking for translation

00:58:29asking to generate

00:58:30image

00:58:31such as what

00:58:31you could do

00:58:32with any LLMs

00:58:33if you go on

00:58:33chat GPT

00:58:34or on Gemini

00:58:34but more than 5%

00:58:36are really

00:58:37direct fact check

00:58:39requests

00:58:39we also find

00:58:40that this

00:58:41small sentence

00:58:42grok

00:58:43is this true

00:58:44has been a real

00:58:45trend

00:58:45with nearly 2%

00:58:47of all tweets

00:58:48mentioning LLMs

00:58:49that are those

00:58:49exact

00:58:50three words

00:58:51to link

00:58:55with the

00:58:56previous presentation

00:58:57we find that

00:58:585% of tweets

00:58:59that are targeted

00:59:01by a fact checking

00:59:02request

00:59:02to an LLM bot

00:59:03so basically

00:59:04someone asking

00:59:05grok

00:59:05is this true

00:59:05after an initial

00:59:07claim

00:59:075% of those tweets

00:59:09have also been

00:59:10targeted by

00:59:11community notes

00:59:12so there's a

00:59:12proposed community

00:59:13notes

00:59:13and more than

00:59:1450%

00:59:15have received

00:59:16at least one

00:59:17community note

00:59:17request

00:59:18so something

00:59:19that we could

00:59:20ask is like

00:59:21is there any

00:59:22like complementarities

00:59:24or substituability

00:59:25between the two

00:59:26systems

00:59:27so for now

00:59:28it's hard to answer

00:59:29this question

00:59:29but on this graph

00:59:31you can see in blue

00:59:32the evolution

00:59:33of the number

00:59:34of tweets

00:59:35on which there's

00:59:36at least one

00:59:36proposed community

00:59:38notes

00:59:38and you can see

00:59:39in red

00:59:40the number of tweets

00:59:42on which there's

00:59:42at least one

00:59:43LLM fact check

00:59:45request

00:59:45to grok

00:59:46or to perplexity

00:59:47and you can see

00:59:48that again

00:59:49it's just pure

00:59:50correlation

00:59:50we are not claiming

00:59:51any causality here

00:59:52but you can see

00:59:53that the point

00:59:54in time

00:59:55where we have

00:59:55a sharp decrease

00:59:57in the number

00:59:57of notes

00:59:58on which there's

00:59:58number of tweets

00:59:59on which there's

01:00:00a community notes

01:00:00really coincide

01:00:02with the launch

01:00:03of grok

01:00:04LLM boat

01:00:04on X

01:00:05which was early

01:00:06March this year

01:00:07after what we do

01:00:11we look at

01:00:11the topics

01:00:12of the claim

01:00:13so in order

01:00:14to do that

01:00:14we have two

01:00:15approach

01:00:15one is using

01:00:17classifier

01:00:18the other one

01:00:19so that's what

01:00:19you can see

01:00:20on the left

01:00:21with primary topic

01:00:22and the other one

01:00:23is using

01:00:24sentence barrier

01:00:25transformer models

01:00:26so topic modeling

01:00:27methods

01:00:27to try to assess

01:00:28on which type

01:00:29of content

01:00:30are people asking

01:00:31for this grok

01:00:32is this true

01:00:33fact check request

01:00:34so we focus

01:00:35on tweets

01:00:35in English

01:00:36and we find

01:00:37that

01:00:38if you look

01:00:39like nearly

01:00:3940% of the tweets

01:00:41are related

01:00:41to politics

01:00:42elections

01:00:42war

01:00:43and geopolitics

01:00:45the other main

01:00:46topics are

01:00:47like economy

01:00:47and finance

01:00:48technology

01:00:49and AI

01:00:49celebrity

01:00:50and entertainment

01:00:50and if you look

01:00:51on the right

01:00:52part

01:00:53so the most

01:00:54frequent topic

01:00:55is soccer

01:00:55so we

01:00:56nearly 3%

01:00:57we have a

01:00:58robustness check

01:00:59on which we

01:00:59remove certain

01:01:00topics

01:01:01but you can see

01:01:02that very quickly

01:01:02it turned to

01:01:03like standard

01:01:04topic

01:01:05on which there's

01:01:05some debates

01:01:06or large number

01:01:07of misinformation

01:01:07including the

01:01:09Russia

01:01:09Ukraine war

01:01:10gender and

01:01:11sexuality issues

01:01:12vaccine and

01:01:12COVID

01:01:12directly about

01:01:14Elon Musk

01:01:15about racial

01:01:16injustice

01:01:17and slavery

01:01:18and on the

01:01:19right

01:01:20it's fully

01:01:21automatic

01:01:22so labels

01:01:23are given

01:01:23by a large

01:01:24language model

01:01:25based on

01:01:25top keywords

01:01:26extracted

01:01:26from a

01:01:27BERT

01:01:28topic model

01:01:30then when we

01:01:31look over time

01:01:32since the beginning

01:01:33of our period

01:01:33so we start

01:01:34in March

01:01:35because those

01:01:35LLM's boats

01:01:36were launched

01:01:37in March

01:01:37this displays

01:01:39the most

01:01:39the trending

01:01:40topics each

01:01:41week

01:01:42and trending

01:01:42topics

01:01:43are identified

01:01:44on the

01:01:45topic

01:01:46on the

01:01:46BERT

01:01:47topic

01:01:47modeling

01:01:47that

01:01:48experienced

01:01:48the largest

01:01:49increase

01:01:50from a week

01:01:50to another

01:01:51and we can

01:01:52see that it's

01:01:52perfectly related

01:01:53to what

01:01:54happened

01:01:54during those

01:01:55last six

01:01:56months

01:01:56we can see

01:01:57some discussion

01:01:58about trade

01:01:59and tariff

01:02:00we can see

01:02:00gang deportation

01:02:01case

01:02:02which was

01:02:02the ICE

01:02:03gang

01:02:04Los Angeles

01:02:04protest

01:02:05we have

01:02:06a plane

01:02:06crash

01:02:07incident

01:02:07related to

01:02:08the Indian

01:02:09plane crash

01:02:10in June

01:02:12recently

01:02:13a lot

01:02:15of tweet

01:02:15about

01:02:16Iran

01:02:16nuclear

01:02:17enrichment

01:02:17Israel

01:02:18Iran

01:02:18ceasefire

01:02:18and we have

01:02:19recently extended

01:02:20the time

01:02:21period up to

01:02:21September

01:02:22and we have

01:02:23like on the

01:02:23last two

01:02:24weeks of

01:02:24September

01:02:25a huge

01:02:26peak on

01:02:27Charlie Kirk

01:02:28assassination

01:02:28in Utah

01:02:29which shows

01:02:30that automatically

01:02:31we really

01:02:32managed to

01:02:33capture the

01:02:34trends

01:02:34in what

01:02:35people are

01:02:35asking

01:02:36fact-checking

01:02:37for

01:02:38then we

01:02:40look at

01:02:40who asked

01:02:41for fact-check

01:02:42so who

01:02:42uses

01:02:43grok

01:02:43is these

01:02:44true

01:02:44patterns

01:02:45so in order

01:02:45to analyze

01:02:46the political

01:02:47leaning

01:02:47we use a

01:02:48method based

01:02:48on the

01:02:49political

01:02:49elites

01:02:50followed

01:02:50by each

01:02:50users

01:02:51that has

01:02:51been

01:02:51developed

01:02:52actually

01:02:52by my

01:02:53two

01:02:53co-authors

01:02:54on the

01:02:54former

01:02:54paper

01:02:55and we

01:02:56find that

01:02:57fact-checking

01:02:57requests

01:02:58are slightly

01:02:59more

01:02:59frequently

01:03:00initiated

01:03:00by

01:03:00Republican

01:03:01affiliated

01:03:01users

01:03:02than by

01:03:02Democrats

01:03:03it's

01:03:04plus

01:03:044%

01:03:05we also

01:03:06check

01:03:06the

01:03:06composition

01:03:07today

01:03:07so we

01:03:08select

01:03:08a

01:03:08random

01:03:08sample

01:03:09of

01:03:09tweets

01:03:09and we

01:03:10still

01:03:10find

01:03:10that

01:03:10even

01:03:11today

01:03:11with

01:03:11our

01:03:11methodology

01:03:12the

01:03:13composition

01:03:14of X

01:03:15is nearly

01:03:16balanced

01:03:16with close

01:03:17to 50%

01:03:18Republicans

01:03:18and 50%

01:03:19Democrats

01:03:21but it was

01:03:22something like

01:03:2270%

01:03:23Democrats

01:03:24and 30%

01:03:25Republicans

01:03:25two years

01:03:26ago

01:03:26so there

01:03:26is a

01:03:27huge

01:03:27change

01:03:27following

01:03:28Elon Musk

01:03:29acquisition

01:03:29but in

01:03:30the US

01:03:31and with

01:03:31our

01:03:31methods

01:03:32to derive

01:03:32political

01:03:33leaning

01:03:33we have

01:03:34still a

01:03:34balanced

01:03:34sample

01:03:35of

01:03:35Republicans

01:03:36and

01:03:36Democrats

01:03:37what we

01:03:38can see

01:03:38is that

01:03:38Grok

01:03:39is used

01:03:39more

01:03:39by

01:03:40Republicans

01:03:40and

01:03:41Perpaclip City

01:03:42is used

01:03:42much more

01:03:43by Democrats

01:03:43than by

01:03:44Republicans

01:03:45so we

01:03:46show that

01:03:46LLMs

01:03:47may themselves

01:03:48be subject

01:03:48to partisan

01:03:49preference

01:03:50with Grok

01:03:50likely influenced

01:03:51by its

01:03:52association

01:03:52with Elon Musk

01:03:53garnering

01:03:54particular support

01:03:55from Republican

01:03:56users

01:03:57in the

01:03:59question

01:03:59on who

01:04:00get fact

01:04:01checked

01:04:01in community

01:04:02notes

01:04:03it has

01:04:03been

01:04:03shown

01:04:03that

01:04:04users

01:04:05tended

01:04:05to

01:04:06fact

01:04:06check

01:04:06out

01:04:06group

01:04:07members

01:04:07more

01:04:08it

01:04:08means

01:04:08that

01:04:08Republicans

01:04:09were

01:04:10more

01:04:10likely

01:04:10to

01:04:10fact

01:04:11check

01:04:11Democrats

01:04:12and vice

01:04:12versa

01:04:13we find

01:04:14something

01:04:14totally

01:04:14different

01:04:15from

01:04:15LLM

01:04:16fact

01:04:16checking

01:04:16so here

01:04:17we find

01:04:18that

01:04:18in

01:04:19all

01:04:19groups

01:04:19Republicans

01:04:20users

01:04:21are

01:04:21targeted

01:04:22for

01:04:22fact

01:04:22checking

01:04:22requests

01:04:23from

01:04:23LLMs

01:04:24much

01:04:24more

01:04:25frequently

01:04:25than

01:04:25Democrats

01:04:26it's

01:04:27true

01:04:27on

01:04:27all

01:04:27users

01:04:27it's

01:04:28true

01:04:28on

01:04:28Democrats

01:04:29but

01:04:29it's

01:04:29the biggest

01:04:31one

01:04:31is from

01:04:31Republicans

01:04:32asking

01:04:33if the

01:04:35content

01:04:35from

01:04:35another

01:04:36Republican

01:04:36is

01:04:37true

01:04:37so

01:04:37it's

01:04:37very

01:04:37different

01:04:38then

01:04:39we

01:04:39map

01:04:39each

01:04:40answer

01:04:40from

01:04:40the

01:04:40LLM

01:04:41to

01:04:41a

01:04:41veracity

01:04:41score

01:04:42because

01:04:42we

01:04:42have

01:04:43an

01:04:43answer

01:04:43from

01:04:43Grok

01:04:44that

01:04:44says

01:04:44yes

01:04:44it's

01:04:44true

01:04:45no

01:04:45it's

01:04:45wrong

01:04:45so

01:04:46we

01:04:47map

01:04:47each

01:04:47answer

01:04:48to

01:04:48a

01:04:48score

01:04:49and

01:04:49what

01:04:50we

01:04:50find

01:04:50is

01:04:50that

01:04:51the

01:04:51veracity

01:04:51score

01:04:52given

01:04:52by

01:04:53the

01:04:53LLM

01:04:53is

01:04:54significantly

01:04:54lower

01:04:55when

01:04:56the

01:04:56initial

01:04:56claim

01:04:57has

01:05:01is

01:05:01assessed

01:05:02by

01:05:02Grok

01:05:02when

01:05:03we

01:05:03take

01:05:03both

01:05:04the

01:05:04veracity

01:05:05when

01:05:05there's

01:05:05a

01:05:05fact

01:05:06check

01:05:06request

01:05:07plus

01:05:07the

01:05:07volume

01:05:07we

01:05:08find

01:05:08that

01:05:08tweets

01:05:09from

01:05:09Republicans

01:05:10are

01:05:10classified

01:05:11as

01:05:11false

01:05:11twice

01:05:12as

01:05:12often

01:05:12as

01:05:13tweets

01:05:13from

01:05:13Democrats

01:05:14even

01:05:14based

01:05:15on

01:05:15Grok's

01:05:15own

01:05:16assessments

01:05:16of

01:05:17what's

01:05:17true

01:05:17or

01:05:17false

01:05:18We

01:05:19also

01:05:20look

01:05:20at

01:05:20the

01:05:21agreement

01:05:21between

01:05:22LLMs

01:05:22because

01:05:23we have

01:05:23some

01:05:23tweets

01:05:23that

01:05:23tag

01:05:24both

01:05:24LLMs

01:05:25directly

01:05:25so

01:05:25we

01:05:25have

01:05:26the

01:05:26exact

01:05:26same

01:05:26claim

01:05:27at

01:05:27the

01:05:27exact

01:05:27same

01:05:28time

01:05:28period

01:05:28and

01:05:29we

01:05:29find

01:05:30that

01:05:30overall

01:05:30there's

01:05:31a

01:05:31quite

01:05:31strong

01:05:32inter

01:05:33agreement

01:05:34LLM

01:05:35score

01:05:35it

01:05:36means

01:05:36that

01:05:36if

01:05:36you

01:05:36have

01:05:36two

01:05:37different

01:05:37LLMs

01:05:38if

01:05:38something

01:05:38is

01:05:38true

01:05:38or

01:05:39false

01:05:39they

01:05:39tend

01:05:40to

01:05:40answer

01:05:40the

01:05:41same

01:05:41thing

01:05:41in

01:05:42only

01:05:437%

01:05:43of

01:05:44the

01:05:44case

01:05:44they

01:05:44diverge

01:05:45by

01:05:45more

01:05:45than

01:05:4550

01:05:46points

01:05:46it

01:05:46means

01:05:47that

01:05:47one

01:05:47says

01:05:47it

01:05:47is

01:05:47true

01:05:48the

01:05:48other

01:05:48one

01:05:48says

01:05:49it

01:05:49false

01:05:49and

01:05:50we

01:05:50end

01:05:50up

01:05:50with

01:05:50a

01:05:51correlation

01:05:51of

01:05:510.57

01:05:52and

01:05:53we

01:05:54also

01:05:54test

01:05:54with

01:05:54a

01:05:55lot

01:05:55of

01:05:55other

01:05:55LLMs

01:05:56outside

01:05:56the

01:05:57platform

01:05:57but

01:05:57including

01:05:58perplexity

01:05:59GPT

01:06:00GRUX3

01:06:00GRUX4

01:06:01and

01:06:01we

01:06:02find

01:06:02that

01:06:02the

01:06:02correlation

01:06:02between

01:06:03LLM

01:06:03answers

01:06:03on

01:06:04fact

01:06:04checking

01:06:04claims

01:06:05is

01:06:06equal

01:06:07to

01:06:070.67

01:06:08which

01:06:09is

01:06:09very

01:06:09close

01:06:10to

01:06:10the

01:06:10correlation

01:06:11between

01:06:12professional

01:06:12fact

01:06:13checkers

01:06:13so

01:06:14it

01:06:14means

01:06:14there's

01:06:14a

01:06:15high

01:06:15level

01:06:15of

01:06:15consistency

01:06:16across

01:06:16LLM

01:06:17in

01:06:18determining

01:06:20and

01:06:22the

01:06:22last

01:06:22thing

01:06:23which

01:06:23might

01:06:23be

01:06:24even

01:06:24the

01:06:24most

01:06:24important

01:06:25thing

01:06:25is

01:06:25that

01:06:25we

01:06:25asked

01:06:26two

01:06:26independent

01:06:26fact

01:06:27checkers

01:06:27to

01:06:27classify

01:06:28a

01:06:28random

01:06:28sample

01:06:28of

01:06:28tweets

01:06:29and

01:06:30fact

01:06:30checkers

01:06:31so

01:06:31those

01:06:31tweets

01:06:31have been

01:06:32classified

01:06:33in

01:06:33real time

01:06:34by

01:06:34both

01:06:34Grok

01:06:34and

01:06:35perplexity

01:06:35so

01:06:35we

01:06:36derive

01:06:36the

01:06:36answer

01:06:37from

01:06:37Grok

01:06:37and

01:06:38perplexity

01:06:38and

01:06:39the

01:06:39fact

01:06:39checker

01:06:40have

01:06:40the

01:06:40exact

01:06:40same

01:06:40information

01:06:41derives

01:06:41a

01:06:42score

01:06:42between

01:06:420

01:06:42and

01:06:43100

01:06:43and

01:06:44then

01:06:44we

01:06:500.5%

01:06:51now

01:06:52which

01:06:52is

01:06:52not

01:06:53that

01:06:53small

01:06:54but

01:06:54of course

01:06:54not

01:06:55that

01:06:55big

01:06:55and

01:06:56if

01:06:56you

01:06:56look

01:06:57at

01:06:57strong

01:06:57disagreement

01:06:58it

01:06:58means

01:06:58when

01:06:59the

01:06:59LLM

01:06:59says

01:06:59that

01:07:00it's

01:07:00false

01:07:00but

01:07:01both

01:07:01fact

01:07:01checkers

01:07:02say

01:07:02that

01:07:02it's

01:07:02true

01:07:03we

01:07:03are

01:07:04on

01:07:04something

01:07:05like

01:07:0510

01:07:05to

01:07:0515

01:07:06percent

01:07:06of

01:07:07claims

01:07:07on

01:07:08which

01:07:08there's

01:07:08a

01:07:08strong

01:07:09disagreement

01:07:09on

01:07:10the

01:07:10veracity

01:07:10between

01:07:11fact

01:07:12checkers

01:07:12and

01:07:13bots

01:07:13one

01:07:14thing

01:07:14is

01:07:15that

01:07:15LLM

01:07:15are

01:07:16generated

01:07:16in real

01:07:20the

01:07:20event

01:07:21plus

01:07:21we

01:07:22have

01:07:22a

01:07:22measurement

01:07:22error

01:07:23so

01:07:23we

01:07:23basically

01:07:23think

01:07:24that

01:07:24it's

01:07:24a

01:07:24lower

01:07:25bond

01:07:25of

01:07:25the

01:07:25true

01:07:26veracity

01:07:27of

01:07:28LLMs

01:07:28so

01:07:29about

01:07:29my

01:07:30quick

01:07:30conclusion

01:07:30so

01:07:31there's

01:07:31a

01:07:31growing

01:07:31demand

01:07:32for

01:07:32large

01:07:32non-graduate

01:07:33model

01:07:33fact

01:07:33checking

01:07:33it's

01:07:34a

01:07:34huge

01:07:35trend

01:07:35on

01:07:35X

01:07:36and

01:07:37it

01:07:37provides

01:07:37fast

01:07:38and

01:07:38broad

01:07:38coverage

01:07:39so

01:07:39it's

01:07:39going

01:07:39to

01:07:39answer

01:07:39to

01:07:39any

01:07:40request

01:07:40so

01:07:40there's

01:07:41no

01:07:41things

01:07:41about

01:07:41is

01:07:42it

01:07:42like

01:07:42politically

01:07:43polarized

01:07:43the

01:07:45LLM

01:07:45will

01:07:45answer

01:07:45to

01:07:46nearly

01:07:46any

01:07:46fact

01:07:47checking

01:07:47request

01:07:47the

01:07:48speed

01:07:49it

01:07:49answers

01:07:49in

01:07:50a

01:07:50minute

01:07:50to

01:07:50anything

01:07:51comes

01:07:51at

01:07:52the

01:07:52expense

01:07:52of

01:07:52transparency

01:07:53we

01:07:54don't

01:07:54know

01:07:54anything

01:07:55about

01:07:55the

01:07:55source

01:07:56logic

01:07:56or

01:07:56uncertainty

01:07:56level

01:07:57with

01:07:57some

01:07:57information

01:07:58about

01:07:58the

01:07:59LLM

01:07:59so

01:07:59the

01:07:59prompt

01:08:00use

01:08:00but

01:08:00not

01:08:00much

01:08:01the

01:08:03accuracy

01:08:03is

01:08:03far

01:08:03from

01:08:03perfect

01:08:04but

01:08:04still

01:08:05what

01:08:05we

01:08:05think

01:08:06that

01:08:06it

01:08:06still

01:08:06shows

01:08:07some

01:08:07potential

01:08:07to

01:08:08identify

01:08:08and

01:08:09correct

01:08:09inaccurate

01:08:10content

01:08:10scale

01:08:10the

01:08:11question

01:08:11is

01:08:1310%

01:08:13too

01:08:14much

01:08:14and

01:08:14is

01:08:14a

01:08:14way

01:08:15to

01:08:15be

01:08:15closer

01:08:16to

01:08:16five

01:08:16or

01:08:16even

01:08:17less

01:08:17and

01:08:18what

01:08:18would

01:08:19be

01:08:19deemed

01:08:19as

01:08:19acceptable

01:08:20and

01:08:20what

01:08:20we

01:08:21are

01:08:21currently

01:08:21studying

01:08:21is

01:08:21that

01:08:21have

01:08:22things

01:08:22changed

01:08:22with

01:08:23the

01:08:23release

01:08:23of

01:08:23workflow

01:08:24thank

01:08:24you

01:08:24thank

01:08:29you

01:08:29Thomas

01:08:30and

01:08:30thank

01:08:30you

01:08:30to

01:08:30the

01:08:31speaker

01:08:31for

01:08:31this

01:08:31very

01:08:32rich

01:08:32entertaining

01:08:33and

01:08:33stimulating

01:08:34session

01:08:34so

01:08:35what

01:08:35we

01:08:36have

01:08:36seen

01:08:36from

01:08:36different

01:08:36angles

01:08:37that

01:08:37how

01:08:38misinformation

01:08:38challenge

01:08:39our

01:08:39democracies

01:08:40and

01:08:40how

01:08:41both

01:08:41institutions

01:08:41and

01:08:42technologies

01:08:42can

01:08:43help

01:08:44to

01:08:44provide

01:08:45answer

01:08:45that's

01:08:47fascinating

01:08:47and

01:08:48I think

01:08:48there will

01:08:49be a lot

01:08:49of questions

01:08:50in the room

01:08:50otherwise

01:08:50I have

01:08:51some

01:08:51how much

01:08:52time do

01:08:52we have

01:08:52for

01:08:53Q&A

01:08:53five

01:08:55minutes

01:08:56let's

01:09:02start

01:09:02the

01:09:03first

01:09:03and

01:09:04then

01:09:04the

01:09:04second

01:09:04question

01:09:13maybe

01:09:13two

01:09:13for

01:09:14Nicolas

01:09:14the

01:09:16main

01:09:16question

01:09:17is

01:09:17whether

01:09:17the

01:09:18treatment

01:09:18that

01:09:19you

01:09:19apply

01:09:19to

01:09:19the

01:09:20people

01:09:21is

01:09:22it

01:09:22not

01:09:22a

01:09:23little

01:09:23bit

01:09:23too

01:09:23light

01:09:23meaning

01:09:24you

01:09:24just

01:09:25show

01:09:25them

01:09:25five

01:09:26fake

01:09:26news

01:09:26and

01:09:28is

01:09:28it

01:09:29enough

01:09:29to

01:09:29maybe

01:09:30have

01:09:31an

01:09:31effect

01:09:32meaning

01:09:32the

01:09:33amount

01:09:34of

01:09:34fake

01:09:34news

01:09:34is

01:09:35it

01:09:35big

01:09:35enough

01:09:36so

01:09:36has

01:09:36to

01:09:37maybe

01:09:37change

01:09:38their

01:09:38vote

01:09:38and

01:09:39the

01:09:39second

01:09:39one

01:09:39is

01:09:40more

01:09:40whether

01:09:41the

01:09:41result

01:09:41you

01:09:41show

01:09:42is

01:09:42robust

01:09:42and

01:09:43how

01:09:43do

01:09:43you

01:09:44choose

01:09:44who

01:09:44is

01:09:45in

01:09:45the

01:09:45center

01:09:45and

01:09:46who

01:09:46is

01:09:46in

01:09:47the

01:09:47left

01:09:47between

01:09:48the

01:09:48candidates

01:09:49and

01:09:49is

01:09:49robust

01:09:50to

01:09:50this

01:09:50choice

01:09:51I

01:09:55fully

01:09:55agree

01:09:55that

01:09:56for

01:09:56any

01:09:56non

01:09:57result

01:09:57the

01:09:58one

01:09:59hypothesis

01:09:59is

01:10:00always

01:10:00that

01:10:00the

01:10:00treatment

01:10:01was

01:10:01not

01:10:01strong

01:10:01enough

01:10:02so

01:10:02with

01:10:04what I

01:10:04can

01:10:04say

01:10:05is

01:10:05that

01:10:05we

01:10:05use

01:10:06the

01:10:06same

01:10:06kind

01:10:07of

01:10:07exposure

01:10:08to

01:10:08fake

01:10:08news

01:10:08treatment

01:10:09as

01:10:09has

01:10:09been

01:10:09used

01:10:09in

01:10:10previous

01:10:10studies

01:10:10in

01:10:10particular

01:10:11the one

01:10:11in France

01:10:11I was

01:10:12mentioning

01:10:12before

01:10:12and

01:10:14in

01:10:15general

01:10:15what we

01:10:16miss

01:10:16with

01:10:16this

01:10:16kind

01:10:17of

01:10:17experiment

01:10:17that

01:10:18might

01:10:18well

01:10:18be

01:10:18important

01:10:19as

01:10:19well

01:10:19I'm

01:10:19not

01:10:20saying

01:10:20that

01:10:20we

01:10:20have

01:10:20closed

01:10:21the

01:10:21question

01:10:21I'm

01:10:21just

01:10:22saying

01:10:22that

01:10:22we

01:10:22try

01:10:22to

01:10:23see

01:10:23what

01:10:23are

01:10:23the

01:10:23effects

01:10:24in

01:10:24way

01:10:24that

01:10:25will

01:10:25be

01:10:25convincing

01:10:26in

01:10:26terms

01:10:26of

01:10:27really

01:10:27the

01:10:28fake

01:10:28news

01:10:28having

01:10:29an

01:10:29effect

01:10:29on

01:10:30voting

01:10:30is

01:10:30the

01:10:31repeated

01:10:31exposure

01:10:32to

01:10:32such

01:10:32fake

01:10:32news

01:10:33and

01:10:33the

01:10:33possible

01:10:34snowball

01:10:34effect

01:10:35it

01:10:35might

01:10:35have

01:10:35on

01:10:36political

01:10:36opinions

01:10:37on

01:10:37voting

01:10:38and

01:10:38with

01:10:38all

01:10:38that

01:10:38I

01:10:39fully

01:10:39agree

01:10:39there

01:10:40is

01:10:40still

01:10:40a lot

01:10:41to do

01:10:41on

01:10:42this

01:10:42regard

01:10:42at

01:10:44least

01:10:44I

01:10:45mean

01:10:45the

01:10:45mere

01:10:46fact

01:10:46of

01:10:47being

01:10:47exposed

01:10:48to

01:10:48news

01:10:49you

01:10:49remember

01:10:49you

01:10:50didn't

01:10:50see

01:10:50before

01:10:51and

01:10:52you

01:10:52believe

01:10:52as

01:10:53true

01:10:53that

01:10:53happens

01:10:54to be

01:10:54populist

01:10:55narrative

01:10:55that

01:10:56has

01:10:56been

01:10:56fact

01:10:56checked

01:10:56as

01:10:57fake

01:10:57news

01:10:58doesn't

01:10:59have

01:11:00any

01:11:00direct

01:11:00consequence

01:11:01despite

01:11:02the

01:11:02many

01:11:02consequences

01:11:03that

01:11:03have

01:11:03been

01:11:03identified

01:11:04using

01:11:04a

01:11:05similar

01:11:05methodology

01:11:05on

01:11:06other

01:11:06topics

01:11:06that

01:11:07I

01:11:07have

01:11:07very

01:11:07quickly

01:11:07mentioned

01:11:08at

01:11:08the

01:11:08beginning

01:11:09of

01:11:09my

01:11:09talk

01:11:09so

01:11:11regarding

01:11:11your

01:11:11second

01:11:12question

01:11:12it's

01:11:13a

01:11:13very

01:11:13good

01:11:13question

01:11:14and

01:11:14in

01:11:15particular

01:11:15the

01:11:16classification

01:11:16that

01:11:17we

01:11:17use

01:11:17does

01:11:18not

01:11:18rely

01:11:18on

01:11:19the

01:11:19political

01:11:19platform

01:11:20that

01:11:20the

01:11:21candidates

01:11:21use

01:11:23for

01:11:24the

01:11:24election

01:11:25it's

01:11:26based

01:11:26on

01:11:27the

01:11:27way

01:11:29voters

01:11:29typically

01:11:30classify

01:11:31the

01:11:31candidates

01:11:32in

01:11:32the

01:11:32left

01:11:33center

01:11:33right

01:11:34spectrum

01:11:35so

01:11:36we

01:11:36have

01:11:36so

01:11:37there

01:11:37is

01:11:37some

01:11:37data

01:11:38for

01:11:38instance

01:11:39that

01:11:40allows

01:11:40to do

01:11:40that

01:11:41and

01:11:41essentially

01:11:41what

01:11:42the center

01:11:42means

01:11:42is not

01:11:43that

01:11:43you

01:11:43are

01:11:44using

01:11:44a

01:11:44centrist

01:11:45platform

01:11:46for

01:11:47the

01:11:47election

01:11:48but

01:11:48that

01:11:48you

01:11:48are

01:11:48really

01:11:49at

01:11:49the

01:11:49middle

01:11:50in

01:11:50the

01:11:50middle

01:11:50as

01:11:50compared

01:11:51to

01:11:51all

01:11:51the

01:11:51other

01:11:51candidates

01:11:52and

01:11:52so

01:11:52if

01:11:53the

01:11:53idea

01:11:53you

01:11:53have

01:11:53in

01:11:53mind

01:11:54is

01:11:54that

01:11:54Macron

01:11:54could

01:11:55be

01:11:56identified

01:11:56as

01:11:57a

01:11:57right-wing

01:11:58candidate

01:11:59in

01:11:59the

01:11:59last

01:12:00election

01:12:00this

01:12:00is

01:12:01something

01:12:01I

01:12:12I

01:12:13think

01:12:13we

01:12:13have

01:12:13a

01:12:13second

01:12:14question

01:12:14there

01:12:14thank you

01:12:19very much

01:12:19for the

01:12:19fascinating

01:12:20panel

01:12:20I'm from

01:12:21the

01:12:21Canadian

01:12:21regulator

01:12:22the

01:12:22equivalent

01:12:23of

01:12:23Ofcom

01:12:23and

01:12:23Ofcom

01:12:24so I

01:12:25want to

01:12:25ask a

01:12:25question

01:12:26for the

01:12:26entire

01:12:26panel

01:12:27given

01:12:28the

01:12:29emergence

01:12:29of

01:12:29Grok

01:12:30given

01:12:30the

01:12:30emergence

01:12:31of

01:12:31Commons

01:12:31given

01:12:32the

01:12:32unsure

01:12:32impacts

01:12:34on

01:12:35political

01:12:35voting

01:12:35and

01:12:36given

01:12:36the

01:12:36importance

01:12:36of

01:12:37media

01:12:37literacy

01:12:38should

01:12:39regulators

01:12:40be

01:12:40attempting

01:12:41to

01:12:41regulate

01:12:42in

01:12:42this

01:12:42space

01:12:43or

01:12:44is

01:12:45it

01:12:45moving

01:12:45too

01:12:45fast

01:12:46for

01:12:46regulators

01:12:46to

01:12:47step

01:12:49into

01:12:49the

01:12:49space

01:12:49I'd

01:12:50be

01:12:50fascinated

01:12:50based

01:12:51on

01:12:51your

01:12:51research

01:12:51whether

01:12:52you

01:12:52think

01:12:52that

01:12:52Autocom

01:12:53Ofcom

01:12:54CRTC

01:12:55in my

01:12:55case

01:12:56should

01:12:56be

01:12:57stepping

01:12:57into

01:12:57the

01:12:57space

01:12:58because

01:12:58of

01:12:58the

01:12:58importance

01:12:59for

01:12:59democratic

01:13:00values

01:13:00or

01:13:01whether

01:13:01it's

01:13:01simply

01:13:02moving

01:13:02too

01:13:02fast

01:13:03to

01:13:03regulate

01:13:03in

01:13:04this

01:13:04space

01:13:04I

01:13:07think

01:13:08is

01:13:08for

01:13:08all

01:13:08of

01:13:08you

01:13:08Maybe

01:13:12about

01:13:13something

01:13:13that's

01:13:13moving

01:13:13really

01:13:14fast

01:13:14is

01:13:14like

01:13:14the

01:13:14use

01:13:15of

01:13:15LLMs

01:13:17so

01:13:18I

01:13:19think

01:13:19what

01:13:20could

01:13:20be

01:13:20important

01:13:21is

01:13:21to

01:13:21have

01:13:22an

01:13:22idea

01:13:22or

01:13:23at

01:13:23least

01:13:23like

01:13:23either

01:13:24to

01:13:24make

01:13:24people

01:13:24aware

01:13:25of

01:13:25how

01:13:26LLMs

01:13:26work

01:13:27what

01:13:27is

01:13:27in

01:13:27the

01:13:27logic

01:13:28and

01:13:29what

01:13:29is

01:13:29a

01:13:29specific

01:13:30prompt

01:13:30design

01:13:31used

01:13:32by

01:13:32the

01:13:32LLM

01:13:33designers

01:13:34to

01:13:34answer

01:13:35to

01:13:35a

01:13:36question

01:13:36on

01:13:36the

01:13:36platform

01:13:37because

01:13:37for

01:13:38the

01:13:38example

01:13:38of

01:13:38grok

01:13:38if

01:13:39you

01:13:39use

01:13:39grok

01:13:40on

01:13:40grok.com

01:13:41you

01:13:42won't

01:13:42have

01:13:42the

01:13:43same

01:13:43answers

01:13:43as

01:13:44if

01:13:44you

01:13:44use

01:13:44a

01:13:44grok

01:13:45boat

01:13:45on

01:13:46X

01:13:46because

01:13:47there's

01:13:47a

01:13:47specific

01:13:48prompt

01:13:49on

01:13:50grok

01:13:51grok

01:13:52that

01:13:52was

01:13:52made

01:13:52public

01:13:53at

01:13:53one

01:13:53time

01:13:53and

01:13:54for

01:13:54example

01:13:54on

01:13:55the

01:13:55grok

01:13:56boat

01:13:56and

01:13:56that's

01:13:56why

01:13:57in

01:13:57july

01:13:58we

01:13:58have

01:13:58this

01:13:59grok

01:14:00praising

01:14:00hitlers

01:14:01and

01:14:01calling

01:14:02himself

01:14:02mecha

01:14:03hitlers

01:14:03and

01:14:03they

01:14:03were

01:14:04deleting

01:14:04all

01:14:04their

01:14:04tweets

01:14:05during

01:14:05three

01:14:05days

01:14:06is

01:14:06that

01:14:06on

01:14:07the

01:14:07fine

01:14:08tuning

01:14:08of

01:14:09the

01:14:09LLM

01:14:09or

01:14:10the

01:14:10prompt

01:14:10used

01:14:11for

01:14:11the

01:14:11LLM

01:14:11for

01:14:12the

01:14:13grok

01:14:13one

01:14:13there's

01:14:13something

01:14:14close to

01:14:15you

01:14:16should

01:14:16not

01:14:17trust

01:14:17traditional

01:14:18media

01:14:18do

01:14:18not

01:14:19hesitate

01:14:19to

01:14:20be

01:14:20politically

01:14:20incorrect

01:14:21so

01:14:22of course

01:14:23the way

01:14:24you

01:14:24design

01:14:24that

01:14:24will

01:14:25have

01:14:25an

01:14:25effect

01:14:26so

01:14:26something

01:14:27like

01:14:27transparency

01:14:28about

01:14:29LLMs

01:14:31I think

01:14:31is

01:14:31important

01:14:32I don't

01:14:32know

01:14:32how

01:14:32it's

01:14:33feasible

01:14:33and

01:14:34after

01:14:34education

01:14:34about

01:14:35the fact

01:14:35that LLMs

01:14:36might

01:14:36be

01:14:36wrong

01:14:36even

01:14:37more

01:14:37for

01:14:37breaking

01:14:37news

01:14:38events

01:14:38and

01:14:38for

01:14:38real-time

01:14:39events

01:14:39and

01:14:40maybe

01:14:41it's

01:14:41not

01:14:41such

01:14:42a bad

01:14:42opinion

01:14:42it might

01:14:43be

01:14:43better

01:14:43than

01:14:44some

01:14:44other

01:14:44random

01:14:44users

01:14:45but

01:14:45it's

01:14:46not

01:14:46the

01:14:46truth

01:14:46so

01:14:47it

01:14:48might

01:14:49be

01:14:49a

01:14:49compliment

01:14:49in

01:14:49some

01:14:49case

01:14:50if

01:14:54I

01:14:55can

01:14:55add

01:14:55to

01:14:55that

01:14:55I

01:14:56don't

01:14:56think

01:14:56the

01:14:57regulator

01:14:57has

01:14:57an

01:14:58interest

01:14:58in

01:14:58entering

01:14:58in

01:14:59the

01:14:59details

01:14:59of

01:15:00each

01:15:00new

01:15:00technology

01:15:01where

01:15:01it

01:15:01should

01:15:01have

01:15:02a

01:15:02clear

01:15:02definition

01:15:03that

01:15:04would

01:15:04allow

01:15:04to

01:15:04ask

01:15:05the

01:15:05question

01:15:05every

01:15:06time

01:15:06so

01:15:06demonstrate

01:15:07to

01:15:08us

01:15:08how

01:15:08this

01:15:09can

01:15:09be

01:15:09independently

01:15:10checked

01:15:11and

01:15:11demonstrate

01:15:12to

01:15:12us

01:15:12the

01:15:13quality

01:15:13of

01:15:13this

01:15:13thing

01:15:13whatever

01:15:14this

01:15:14thing

01:15:14is

01:15:14and

01:15:15one

01:15:16of

01:15:16the

01:15:16results

01:15:17that

01:15:17we've

01:15:17been

01:15:17showing

01:15:17is

01:15:18that

01:15:19there

01:15:20are

01:15:20different

01:15:20ways

01:15:21to

01:15:21tackle

01:15:22these

01:15:22different

01:15:22technologies

01:15:23they have

01:15:23their

01:15:23particularities

01:15:24but

01:15:24these

01:15:24are

01:15:24the

01:15:25guidelines

01:15:25that

01:15:25I

01:15:25think

01:15:26make

01:15:27sense

01:15:27right

01:15:27so

01:15:27when

01:15:28are

01:15:29we

01:15:32cutting

01:15:32through

01:15:33the

01:15:33noise

01:15:33of

01:15:33what

01:15:33these

01:15:34things

01:15:34are

01:15:35supposed

01:15:35to

01:15:35be

01:15:35doing

01:15:35according

01:15:36to

01:15:36platforms

01:15:36I

01:15:39think

01:15:40it's

01:15:40time

01:15:40to

01:15:41stop

01:15:42let me

01:15:44conclude

01:15:44listening

01:15:47to you

01:15:48and I

01:15:48find

01:15:48this

01:15:48discussion

01:15:50and also

01:15:50the question

01:15:51we raised

01:15:51but also

01:15:52the talk

01:15:52really

01:15:52fascinating

01:15:53and

01:15:53as an

01:15:55economist

01:15:55something

01:15:56came to

01:15:56my mind

01:15:57which I

01:15:57think

01:15:58we cannot

01:15:59stay on

01:16:00that

01:16:00because it

01:16:00will take

01:16:00hours

01:16:01of course

01:16:01but

01:16:01three

01:16:04key words

01:16:04would be

01:16:05innovation

01:16:06transparency

01:16:07and

01:16:07incentives

01:16:08so

01:16:09that

01:16:10could be

01:16:10a research

01:16:10agenda

01:16:11by itself

01:16:11like

01:16:12how

01:16:12can

01:16:13democratic

01:16:13societies

01:16:14encourage

01:16:14innovation

01:16:15and create

01:16:16the right

01:16:16incentives

01:16:17so that

01:16:18people

01:16:18can

01:16:18themselves

01:16:19more easily

01:16:20check

01:16:20information

01:16:21and then

01:16:22lead to

01:16:22a

01:16:22public

01:16:23debate

01:16:23to be

01:16:24more

01:16:24transparent

01:16:24I think

01:16:25this is

01:16:25a bit

01:16:26transversal

01:16:26to all

01:16:26that you

01:16:27proposed

01:16:27and

01:16:27maybe we

01:16:29can discuss

01:16:29that later

01:16:30altogether

01:16:30I don't

01:16:31know

01:16:31but as I

01:16:31said

01:16:31there

01:16:32would be

01:16:32hours of

01:16:33discussion

01:16:33and probably

01:16:3420 papers

01:16:35to write

01:16:35so maybe

01:16:36we can all

01:16:36talk about

01:16:37that during

01:16:38the lunch

01:16:38because now

01:16:38it's lunch

01:16:39and so

01:16:39we all have

01:16:40to move on

01:16:41for the lunch

01:16:41quickly

01:16:42because we're

01:16:42late

01:16:43and so

01:16:54thank you

01:16:54and so

01:16:55thank you

01:16:57you

Catégorie

Transcription

Recommandations