cover of episode Network Manipulation

Network Manipulation

2025/4/30
logo of podcast Data Skeptic

Data Skeptic

AI Deep Dive Transcript
People
M
Manita Pote
Topics
Manita Pote: 我主要研究在线信任与安全,重点关注社交媒体上的协调滥用行为,例如协调回复攻击,这是一种通过协调评论来影响公众舆论或骚扰特定目标的策略。我的研究包括识别这些攻击的目标(例如记者、政治人物)、开发检测这些攻击的机器学习框架(包括预测哪些推文可能成为目标以及哪些用户参与其中),以及分析数据结构(例如树状结构和星状结构)的挑战。我们使用推文级和回复级特征,包括回复之间的相似性,来构建模型。模型评估指标包括精确率、召回率和AUC分数。虽然模型在某些活动中表现良好,但在其他活动中表现不佳,这可能与不同活动中使用的策略有关。 此外,我还研究了数据删除在社交媒体操纵中的作用。我们发现,通过删除推文,恶意行为者可以规避API限制、操纵算法以及隐藏其活动。这项研究利用合规数据流来估计删除行为的规模,发现API数据会低估大约45%的删除行为。删除行为的动机可能包括商品促销、粉丝操纵和垃圾邮件活动。 目前,我的研究重点是第三方应用程序如何被用于协调操纵活动,这与Cambridge Analytica事件类似,但视角不同。我关注的是这些活动的物流和运营方面,而不是账户行为。生成式AI的使用使得检测协调操纵活动变得更加困难,这给研究人员带来了新的挑战。 Kyle: 在与Manita Pote的访谈中,我们探讨了社交媒体上协调操纵的复杂性,特别是协调回复攻击。这些攻击通常针对有影响力的人物,例如记者和政治家。Manita Pote的研究重点是利用机器学习模型检测这些活动,该模型使用结构和行为特征来识别参与者和受攻击的推文。此外,我们还讨论了数据删除在规避审核和操纵参与度指标中的作用,以及学术界与平台之间在数据访问和隐私保护方面的平衡问题。Manita Pote的研究为应对社交媒体上的协调操纵提供了宝贵的框架,并突出了生成式AI带来的新挑战。

Deep Dive

Shownotes Transcript

Translations:
中文

You're listening to Data Skeptic: Graphs and Networks, the podcast exploring how the graph data structure has an impact in science, industry, and elsewhere. On the last episode, listeners would have heard we were talking about the new course that you're teaching.

So Asaf, can you give us any update on the course? How's it going? Well, in the beginning of each class, I usually do some kind of like a Kahoot or a Mentimeter. You know, that's like a questionnaire for everyone to answer. And you can see the answers on the board. Cool. So you can understand if people really understood the material or not.

When everything goes well, it's a long-term distribution. You get the one answer right and all the wrong answers, you know, flat. So lately I'm getting, let's say, an even distribution, so I need to check what I'm doing wrong. What's your goal? What's the optimal distribution you'd like to achieve?

I'd like to achieve, you know, the Paolo, not the Paolo, of course, the long tail distribution. I want one column that has the right answer and all the other columns, you know, with zero wrong answers. I think that's great, actually. I've always thought, especially as I give lectures and things like that,

I wish there was a real-time feedback. You can kind of watch people's faces. There's all of the humanistic traits that have always been there, but those aren't always very telling either. If there were sort of a real-time anonymous measurement of some kind or who actually understands what I'm saying right now, who's following the proof?

Exactly. Looks can be deceiving. You know, people nod their heads and they don't understand what you're saying. So actually, there are two methods. One, after I explained the long tail principle, I asked people how many friends do they have on Facebook? And usually the answers are between 200 or 2,000.

If the answers are between 200 and 2,000, it seems logical to extrapolate that most people have between 200 and 2,000 friends. And they nod again. And then I understand they didn't get it. Okay, so then I say, well, it's wrong because there are just a few people, a few people. No, it's 1% from...

three and a half billion MAUs, monthly active users. But still, it's, let's say, 1% with thousands of friends and most of the users have just a few friends. We don't know them because they just have a few friends. But that's one way to sanity check.

if anyone understands. Yeah, I like your technique. You offer something seemingly intuitive and you see if they nod and fail to see that it's actually deceptively counterintuitive. When you think about network laws, it's so unintuitive because you feel like we're human beings, we have free will, we can do whatever we want.

But still, without talking about it, physical laws that govern society and make us behave in ways that we didn't plan to behave. Nobody said we need one giant city, one metropole, and lots of small villages. Nobody said it. But that's how we organize. Nobody said you need a few influencers and everyone else, you're allowed just to be friends with just two or three other users.

But still, we behave the same. I think the most interesting part is like when you look at Twitter data. In Facebook, you can say it's undirected connection because you're my friend, I'm your friend, and you can say who initiated it. But still, but okay, it's...

In Twitter, it's obvious, right? Those followers, when you aggregate the graph, you can see that if you look at the distribution of the followers and the following, it's the same distribution. When you say, who do I follow? It's my decision and my decision alone, right? I decide how many people I follow. But who follows me? It's not just one person's decision, right? It's the decision of the crowd.

The distribution of followers on Twitter and following is the same distribution. That means that people behave on the personal level and on the level of society the same. So people were, after taking your course, less inclined to believe in free will? No, because they know I grade them.

What free will? Fair point. Well, I'm glad you brought up Twitter as an example, because we're going to get into that in today's interview. Twitter is a good example of a social network. There are many others, but Twitter is the one that our guest today, Manita, studied. There's a couple of interesting points about her work, but broadly speaking, she's looking at

Like when people kind of gang up and have a coordinated crowd attack against someone. A few years ago, it was a problem to actually measure what are the effects of, let's say, coordinated bot attacks on Twitter or other social media.

In recent years, I think maybe one or two, with the help of Elon Musk, he did something right, at least for this end, and you can use the view count, right, of every tweet. So when...

There's a malicious intent to manipulate the discourse on subject. Let's say there's a person that really cares about misinformation. He really wants to misinform people. So if he will just use bots that use the same pictures or same text, the machine learning algorithms will get him. Right. It's so simple.

It makes the network analysis more important than just looking at the users one by one by using machine learning techniques. You can find them all at once by using network analysis because you can see who follows whom and aggregate the clusters, the communities. You can see the communities. This is actually an interesting case I got. I had when we found a bot farm that had two kinds of users, exactly this case. So the users that posted comments and followed no one

and users that follow the users that posted the comments. Let's call them the boosters. It's a classic case of page rank manipulation. When they follow the users that comment, they give them some score, right? There's some page rank score. And because the users that post the comment don't follow anyone, they keep the score to themselves. If you use bots to boost your followers, like I don't mean like 200, 300, like they used on this bot farm, but

let's say you want 10,000 followers and you use the bot farms to do it, it will immediately sound all the alarms on the platforms. Okay, so I guess there's a threshold. So you need...

actual cubs to follow you. Well, you certainly outlined a good form of manipulation that goes on in social media. And our interview today hopefully proposes a good framework for combating or at least detecting that in a two-stage fashion. First, they detect the types of tweets that might get these coordinated replies. Because I suppose if you're like a politician, you're just sharing all the time saying, here's what I had for breakfast. It's not very controversial, but

But if you say vote yes on Proposition 1, that would be a good reason that a coordinated attack might happen. So they're able to recognize some of those tweets. And then secondarily, the question becomes, okay, it got a lot of responses. Is that because the public is enthusiastic? Or is that because there's a coordinated attack here? And if so, which are the users that are genuine? And which are those that are involved in that attack? So we get into the framework that approaches that problem today in the interview. ♪

My name is Manita Pote. I'm a PhD student here in Indiana University, Uplomington, and I mainly research on online trust and safety with focus on coordinated abuse and social media. Can you share a few details on what online trust and safety covers?

So online trust and safety basically means the users are having a good experience. There is no any kind of manipulation or like bad experiences. Some of the bad experience could be exposure to the content that you did not want it to. Here lies the content moderation aspect.

Sometimes unknowingly in your feed there are content that you didn't want to see because some other party are manipulating to push their content. So that could be one area of trust and safety. Sometimes there could be cases of your accounts being hacked or used for malicious purposes without your consent, where the user privacy policy protection comes into it.

These days, there are aspects of coordinated manipulation, especially in the realm of political campaigns. So it's trying to influence your behavior or your perception of the world, but it's not the reality. So it's the people trying to create this

different alternate reality that is not true. So that is also one kind of trust and safety aspect of it. So there are various kinds. With regard to content moderation, I happen to not like country music, but I think you mean something a little more specific than that. I don't think it lies in the area of content moderation.

Since you don't like it, basically the algorithm would be able to pick it up based on your past behavior. In case of content moderation, there could be a case of, for example, a pornography video in the feed, right? That is considered to be not within the realm of policy of the platform itself. There could be like a hatred comment against somebody or towards you as well. So this is also not within the realm of platform policy, what is allowed, what is not allowed.

So in those cases, things that are against platform policy that has to be moderated, you make sure that that's not happening in the platform. That's where I think content moderation lies. What platforms are interesting to you from a research perspective? So far we have been studying Twitter, or now it's called X. I think the main reason for that was the data availability.

It was easier to get data for Twitter that had academic access. You could get either a 1% random sample, 10% random sample, or all of the Twitter if you had that access. You could ask for that access as a researcher. But now that has become far less because of a lot of changes in the company's policy. So far, I have worked on Twitter, but now people are more shifting towards the blue sky.

because the data access is easier. And I have some vague understanding that it's gotten harder to get the data. Back in the day, there was a nice API I played around with a little bit. It was perfect from a developer's point of view. I could get whatever I wanted. What's the current state? I know it's less than that, but is there anything available to you? Yes, we do. But I think for a researcher, it's not enough. Like 100 tweets per month, that is far less sample than what actually is.

So I don't know if you could make a good conclusion from that amount of data.

And it's also expensive as well. There are different tiers, like 5K for some access, like 10K for access. I think from a researcher's perspective, that is not feasible. That amount of expenses is not feasible unless somebody is funding you. So yeah, it has become much more harder in terms of getting data for studies. Yeah, either people rely on the data they have already had in the past, past data. So for example, our research lab has the data

The data they used to get like 10% sample. Now, if somebody wants to study something, they could utilize that data. But getting new data is much harder. And what is a coordinated reply attack? Before we dive into this, we need to understand a little bit more of a context of this research. This research was specifically done in the area of influence campaign.

So, influence campaign or information campaign. These campaigns are coordinated ways of manipulating your opinion or influencing your behavior.

These campaigns especially happen in the political case during elections or some kind of other campaign where people are trying to push some kind of agenda. One of the strategies using those coordinated campaigns is to harass somebody or like dogpiling, we call it, like put some negative comments on somebody.

or trying to artificially boost somebody by putting on good comments about somebody. Like, for example, if I want to boost some leader or some representative from my district or something, I would put, oh, this guy is great. Like, this guy is doing great. So everybody duckpiling on his comment section, putting positive comments, either in the form of,

but that is not a natural or organic support. What if I was part of some large group that were, I don't know, a political action group, and we got together and all went at the same time and downvoted the opponent's point of view? Is that what we're discussing? So that would be one way of coordinating the downvoting action. The paper that we just talked about was about coordinating a way of putting comments in somebody's post. So like harassment or an argument

and organic support. We are trying to look into the campaigns that employ this strategy for boosting the person or harassed person. So this study has a lot of components to it. The first part is the case study that we did to look into who are the people these campaign, the people who are involved in campaign are targeting. So for example, a journalist, media personality, political parties, where the prominent targets of these campaigns.

So that was one finding of this paper. The second part is to propose how can we detect these people who are involved in such kind of coordinated attacks. Our idea was to present a framework based on the people they engage with itself, the target itself. We look into the tweets from the targets that were attacked.

and the tweets that were not attacked. So when we compare those positive class and negative class in case of detection, if we put together, can we differentiate the tweets that get targeted and based on the replies on those tweets, can we find those accounts?

So it's a way to find accounts involving coordinated attack based on the target. Well, it would be novel to be able to predict which tweet is going to be a target. What methodology do you use to go about it? We propose like a framework of two machine learning models. One first would predict whether the tweets get coordinated reply or not.

And based on the prediction of this model, we will look into the features from the people who leave the comments. And then second machine learning model would predict whether a user or a replier is involved or not.

So first we'll predict whether a tweet got coordinated by or not. Second model will predict whether an account was involved in that attack or not. So it's a two part framework. We did our own feature engineering, like it's all done by a person. Like feature engineering part I think is the most interesting part of this whole framework. Could we dive deep there then? What are the either specific features or broad categories of features you're looking at?

Everyone's talking about AI these days, right? It's changing how we work, how we learn, and how we interact with the world at a tremendous pace. It's a gold rush at the frontier. But if we're not careful, we might end up in a heap of trouble.

That's why I recommend you check out Red Hat's podcast, Compiler. This season on Compiler, they're diving deep into how AI is reshaping the world we live in. From the ethics of automation to the code behind machine learning, it's breaking down the requirements, capabilities, and implications of using AI. Check out the new season of Compiler, an original podcast from Red Hat. Subscribe now wherever you get podcasts. ♪

Either specific features or broad categories of features you're looking at. When we think of replies to a tweet, then it usually is a tree structure. In the sense of graphs, a tree structure, right? So a tweet has a reply. A reply can also have a reply. So it's a tree structure. So how to go about even doing a feature engineering of this? It's a hard task.

Not only that, in our case, we had a very two different nature of data set as well. The campaign data that we had had more like a star structure. So we were not able to construct a whole tree structure from the reply. We knew which user commented on which tweet, but we were not able to construct the whole tree. So it was more like a star structure.

In the control dataset that we constructed, the tweets that didn't get any reply from the campaign people were more like a tree structure. So it's two different data structures to balance it out so that we are working with the same kind of structure. We converted the tree into star structure itself. So we have a tweet that has a bunch of replies. Now we need to create a feature for this.

We mainly use tweet-level features and the reply-level features. The tweet-level feature has individual numerical value for each tweet, but in case of reply-level feature, it is a distribution, right? So even if you look at the engagement, each reply can have each engagement metrics, like how many likes each reply got.

So for each tweet, the reply level features are all distributional values. So we try to convert these distributional values into a summary statistics for each tweet. So it's a tweet level feature and the reply level feature. The reply level feature has all summary statistics based on the reply level attributes like maybe engagement, maybe how much

how much like each reply got. So we aggregated those into one tabular form. What about text-based natural language processing features? There are different features that could be used. One is you could have your sentiment score as well, right? Like one could be hatred, one like, you know, toxicity scores like that. But one thing toxicity score are like positive scores sometimes.

some score related to sentiment. We could have done that too, but we didn't do it because in a campaign, these comments could be supporting comments as well as some kind of harassment case too. So we didn't have that distinction in our tweets. So we refrain from using that score.

Instead, we just use how similar the replies were. For each tweet and its replies, we calculate the pairwise similarity between the replies, so how similar the replies were. And we use the summary statics of that distribution of similarity, since a tweet can have many pair of reply similarity.

How do you distinguish between a coordinated attack and a just broadly unpopular thing, someone might say, where they get a ton of vitriol because it's a spicy take? This is like kind of trying to distinguish between what is authentic and what is not authentic kind of question, right? Which is very hard to operationalize. What is it? In this case...

We are assuming that these are based on a malicious intent because the data itself came from a campaign and the sheer number of comments in the post itself that are very, very highly similar. So that is one way to say it's actually coordinated and this is not. Would it be correct to assume that the coordinated attacks are probably botnets?

Could be botnet, maybe not as well because these campaigns used to be easier to detect and more like bot-like behavior. Bot has like, they are posting at the same time at the very minimal time interval. But these days, there have been more difficult to say that these are actual botnets. The time doesn't synchronize, but if you look at their reply level, then they're exactly the same. They could be a mix of bots and users as well.

Yeah, it could be, but we cannot with exact say that these are actually bots only. Well, when you look at the overall framework, I don't know if you have exactly any access to ground truth or not. How do you evaluate the quality of the output? This is a very novel task and definitely we inferred like some kind of control and the positive case. The campaign data that we were provided was from Twitter itself. Like these are the malicious messages.

campaign data that you can use in any way. So in a way it is a ground truth, in my opinion.

But this is a very novel task and we don't have anything to compare like what people have done in the past and how we improved. So in that sense, yes, the baseline comparison is not there, but we try to compare our results with different other campaigns as well. So yeah, we trade on one and try to see how it generalizes over other different campaigns for both of the models. But if we trust your labels that you use as ground truth in training the classifier and whatnot, let's just say those are for sure right now.

Could you share some details on the metrics you look at to assess the quality of the model?

Yes, we looked at effort score, precision recall as well, AUC score as well. We tried to experiment every parameter that we define in the beginning of our task. Or five or more replies, which is very arbitrary. We try to change that reply to what happens if it is 10? What happens if it is 15? What happens if it is 20? So we looked at that parameter as well. We also looked into the data ratio. Okay, what happens if it's extremely imbalanced? Will it be able to

proper work or not? What happens if we train on one campaign and test on other campaigns? Will it generalize or not? So we did every kind of test we could think of or like we assumed as a parameter in the beginning. Well, maybe we could also touch on the second part of the framework where users or have we just moved into that where users and the replyers are involved or not?

The first part was classifying tweets, but the second part is more on classifying a user or a replier that was involved in coordinated reply attack. The first part would give us a tweet that it potentially got coordinated reply,

Then based on the replyers involved in that tweet, we would create another set of features that would be used in the second part of machine learning model. These are based on replyers metadata, like how old these accounts are, how often they involve in attacking, like commenting the replyers.

and how fast they were in replying, how many followers following they had. So it's a metadata, reply metadata level, and also the replies level, like what reply they put in. So we constructed a feature based on that and used that feature for the second part of the machine learning model. And roughly speaking then, how would you grade the model? I don't know if it matters exactly what the AOC was or whatever, but...

Is this something that's read that's... To what degree is it doing its job? So there were some campaigns that it worked very well, but in some campaigns it didn't work well. So for example, there was one case of Egypt, their scores were very, very low.

But we have to also understand that these kinds of tactics exist in all of the campaigns, right? It might exist, it might not exist. So in those campaigns that didn't work well, probably these strategies were not used as a main strategy. So that could be a possible explanation for why this model didn't work for some of the campaigns.

performed well for some of the campaigns like Serbia and based on the Stanford reports as well, the strategy we just discussed actually happened in that campaign. So it performed well on that case. And there was, I think, Turkey that also worked well in Turkey. For Egypt, it didn't work well because probably it was not the main strategy being used in that campaign. So yeah, it did work in some case. It didn't work in some case.

In terms of next steps, obviously there could be more research to look into the cases where it was not as effective as others. But do you see a broader goal here? Is this something that's, I don't know if there's a commercial angle or something you publish? Where does it go? The main idea that started this project was how to find these campaign accounts, the accounts that are involved in campaign campaigns.

It depends on the context of a campaign. It depends on the political context of a country and it's very different for different countries, right? So if you don't have any political understanding, contextual understanding, how would you even go about finding this accounts involvement campaign? So we try to look into the common behavior

of these accounts involvement. That's where the target came in. Who were they involving in mass amount? Our one analysis found out that it's basically the influential people they were involved in, like the people who have higher number of following. Then the natural question was, can we find these accounts just by tracking these influential people?

So, you know, okay, journalists, if somebody has labeled their account as journalist, can we just keep on tracking those people to find these accounts? That was the main idea. So if you, we have to think of in a commercial way, the platform could use it as a way to track influential people to find the campaign accounts, especially coordinated campaign accounts. That was the main motive of this whole project.

Yeah, someone like a journalist is sort of inherently a honeypot in a certain sense, right? Yes, yes. A honeypot is the right word. So once you've done, I guess, the reverse honeypot technique and found the coordinated group,

What's to be done? So though the models predict that they are involved, there has to be a manual. Somebody has to look at them manually to even flag them or suspend them. So that manual inspection comes into play, has to come into play anyway.

So a platform can just suspend them for their activity or like even make them delete their account. So it's more about suspension of the accounts that are involved in such attacks. So make it more safer for the people, more enjoyable experience for people to use social media platforms. Basically, the commercial aspect would be that.

So these platforms, I mean, we could have our own opinions about if they're doing a good job or a bad job, but they're doing some effort trying to filter out bad actors like this. And those bad actors are responding. I know from one of your other papers, manipulating Twitter through deletions, people have found a clever way to go under the radar a bit. Can you share some details on this work?

This is a very interesting work from our lab, I think, because so far all the research has been done based on the data they get from API, but the data that gets missed is the one that has been deleted, right? We have no any point to study about it, like make an inference about it.

So this whole paper was trying to see how much we underestimate our analysis when we just look into the data from API and further what kind of malicious activities that goes into a platform in form of deletion.

So, you know, people post in high quantity and delete it. No traces left behind. Researchers never get that data anyway. So we know it happens, but how to study this? What is the underestimation that we make when we study just based on API?

There are many different components to this paper. The first part was to see how much we underestimate when we study data just based on the API, how much we miss. The second part is what kind of manipulation is happening in the form of deletion.

I think we find one on the API limit, circumvention of API limit, and the second part of the manipulation is coordinated like and unlike pattern to manipulate the algorithm. So I think these three components, underestimation, circumvention of the API limit, and third,

coordinated like and unlike with the main different findings of this paper. Yeah, these three behaviors I had never thought of, heard about, or considered before reading your paper. How did you encounter these in the first place, make this discovery? People have, like, there are research people have done that, like,

For example, one paper was on ephemeralized surfing. This was one of our past postdocs' work. In this work, he finds that people would post hashtags a lot in the first few seconds of the post just to manipulate that hashtag ranking and later delete those hashtags or tweets with those hashtags. So it's a way to manipulate those trending topics.

it shows that the deletion really is a like a manipulating strategy that people employ so people have mentioned it in the past but there was not a clear way to study it so our lab started working on this based on the compliance firehose data that we had so compliance firehose data is

a notification that we get from our Twitter every time somebody deletes something. This notification is for us to delete the data that we have from a user on our part. So this compliance firehose is not available to everybody, but it's available to the researcher who keeps the data. So our lab happens to keep that data. So one way of inferring how much deletion happened was to look into the metadata and into compliance firehose.

We cannot look at the content that would be the violation of terms of service, but we could look at the metadata like how many deletions that happen, keeping in mind that we are not exposing people, you know, identity or something. We looked into this compliance firehose and looked into how much each account does deletion in a day.

We took a threshold of 10 or more for our analysis. Since it's a whole of a Twitter, it's a lot of data. We have to remember that this is a whole of a Twitter. We cannot look at every account. So we filtered based on 10 or more deletion in a day. So this way we got the accounts that delete more than 10 tweets in a day from our compliance firehose.

Then we compared how much this user deleted based on the API request. So every day we would parse the compliance firewalls to get actual deletion. And then for each user, we would again request an API to see how much tweet count has decreased in the following day to just make an inference about how much they deleted. So we compare the actual deletion with inferred deletion from the API and make an estimate. So how much

are we underestimating our analysis based on the comparison? We underestimate around 45% based on what we get from our API and when we compare it, the actual deletion that happens. So that was one finding. Roughly speaking, how prevalent is this phenomenon? That is also very hard to estimate, like how prevalent it is. So I don't know how prevalent it is, but it does exist in terms of the past research that people have done

done. For example, in one of the researchers, our past researchers,

work was on Twitter trains. So people would post mentioning others to follow like and the others would start doing the same. So it would create a network of people following each other, boosting each other. And later they would delete these posts so that they don't have any evidence of this happening later. So it exists in different, different forms, but we don't know actually how much this happens, how much in what amount this happens.

To what end are they doing this? Do you have any insight as to... Obviously, it's some of the things you mentioned, like rank manipulation and whatnot, but does it tend to be political or commercial, or is there any themes throughout it? Again, going back to my previous comment, we cannot look into what was deleted. Based on the deleted data, we cannot say what were the topics, but...

Looking at the profiles of some of the accounts that existed after our analysis, they were more like merchandise, follower manipulation kind of accounts. Oh, if you follow 10 accounts, I will follow other 10 more. Those kind of accounts. Some were posting cryptic tweets that didn't have more meaning to it. So we could not place it in a particular political realm. It seems more like a spam kind of activity.

And I know this is a little bit out of scope, but do you have any sense of the degree to which these people are able to influence or maybe annoy one of their targets or change perspectives? That's a very hard question to even study.

I don't know how would one go about it because it comes into play of algorithms as well, right? How does the algorithm would show a content based on the influence or the manipulation of these actors? How that would affect what is shown to the people? This is a very hard question and I think I don't have a clear estimate or understanding of how would one go about studying the effect or impact of this manipulation. Yeah, I'm not sure anyone does, but...

It's certainly a lot of activity trying to do it, so they must be having some effect you would assume, right? Yes. In case of, so for example, in case of this like and unlike, it has some kind of financial aspect to it too, right? So like popularity, number of views, these do have at the end some ties to financial aspect of it. So probably this manipulation do has an effect.

Do you find that there are some platforms that are more targeted? I know you have mostly studied Twitter, but are these phenomenon also in Blue Sky and other places? I don't know about the Blue Sky since it has been a very newer platform and I am not sure whether the influence operation or information operation has been studied in case of Blue Sky. But there are like reports from Twitter, Facebook and TikTok, like they release every year about how they found this influence campaign.

What were the actors doing in this campaigns? They do mention like, oh, these people were targeting this person or this individual. They were trying to do this in this way. So yeah, there is like a lot of reports on this in the different, different platforms, actually. TikTok, I know TikTok and Facebook, they release every year, every few years.

Well, Facebook, this was a while ago, but they were highly criticized for this Cambridge Analytica thing that the API was too open, I guess, so Cambridge was able to do things most users would prefer they didn't. In a way, that implies maybe Facebook should close the door, but that's unideal for researchers like yourself. Do you have any thoughts, having worked at least directly with Twitter, what's an ideal way for that relationship to be developed between academia and a platform?

As far as I know of Cambridge Analytica, I think that happened because the third-party apps, like the web apps that were running on top of Facebook, sold the data of the user to Cambridge Analytica.

So they're like indirectly the Facebook comes into play, like they didn't have that check or like intervention. Definitely the getting data from Facebook is much harder and the kind of openness we had with Twitter is not with Facebook. I think there's always a scrutiny that goes into play whenever any researcher requests data from Facebook. So I don't

I don't know how that balance of, you know, it's open to researchers, but yes, we are also conscious about not having the same incident as before, you know, for Facebook or any of the platform. For example, in case of TikTok, also getting data, the kind of data that we want is harder to get. They do provide it, but I don't know the usefulness of it.

In conclusion, yes, they are trying to make it available to researchers, but I don't know how much that is useful from the research perspective. So I don't know the balance in between how much research should get access, how much they should not, and what is the company's policy, their protection itself. So I absolutely have no idea about that. A difficult balancing act for sure. What's next for you in your work?

I actually am working on another project that is looking at how third-party app manipulation happens. So this is very similar to Cambridge Analytica kind of analysis, but from a little different perspective, where I'm trying to see how unknown apps are trying to coordinate, run these coordinated campaigns.

So I'm trying to look more from a logistical and operational aspect of these campaigns rather than the account behavior. So one of the findings from this paper is that these accounts are run from the same app, like same unknown apps. So rather than looking at the behavioral aspect of it, I think it is better to look from like the app aspect or operational aspect of these campaigns. Then you can find more coordinated campaigns based on this rather than just looking at the behavioral aspect. So that is my work in progress.

Now, I think the campaigns have shifted more into like generative AI case, use of, you know, fake profiles, fake texts, like generative AI texts. They are more influencing, more convincing. And even like accounts, fake accounts that act more like human, it's,

they don't have as similar behavioral traits as it used to be in the past so it becomes more harder to detect more to tell which is right which is fake which is not fake so it is going in more and more like harder to understand in that direction so yeah there are a lot of works to be done in this direction in my opinion and it is a challenge for researchers as well like a

how generated AI has shifted everything. Yeah, it's certainly much easier to generate a convincing profile than it was many years ago using AI. That's making your work harder, I'm sure. Is it a winnable battle? How pessimistic or optimistic are you about the future? Yeah, I think being pessimistic is the easiest way. But like, you know, it's just way to think. But I think there haven't been very innovative ideas in the research community.

as well. I don't know the full detail of it, but one idea that I heard about or just happened to know about was a similar idea to the blockchain. I don't know how it works, but how we can use a similar concept of blockchain in posting content, like trusting parties. So I think researchers have been coming up with very, very new and novel ideas. If there is a problem, there will always be a certain solution. So I think I'm more on a pessimistic side,

Yeah, and I think in terms of regulation from the government and AI policy, people have been working in that direction too. So I think

There are hope for it. We cannot be pessimistic. And you're probably not at a point where everything's planned out yet, but do you have any thoughts on where you'll go after graduation? I'm still a little in a dilemma about like whether I want to go for, you know, academic route or more on a think tank route, you know, more on a policy side or a academia industry itself. So I'm still figuring that out and see where I fit more. Like, you know, I have to find my area as well, like niche area.

So, yes, academia is one way, one place. Policy think tanks would be one way. Industry also does a lot of work in this area. Like Facebook has their own team that works on insurance operations. TikTok has it. I have not yet pinpointed one direction, but I'm open to all direction where I fit. Cool. A lot of opportunity for sure.

Is there anywhere listeners can follow you online? Yes, I'm on Twitter. My name is Manita Pote, which is with no space, where they can follow me. I'm also in a Google Scholar. Yes. All right. We'll have links in the show notes for people to follow up. Manita, thank you so much for taking the time to come on and share your work. Yeah.

Thank you, Kyle, for this opportunity. This is my very first time and I was excited about it, you know, and also nervous. You have been doing this for a very long time. You know, a lot of things. Thank you so much. This was great. Thank you, Kyle. Thank you for the opportunity.