UBNow

Thomas Dinki: Yini Zhang was in college when social media started taking off, and she was optimistic about its potential.

Yini Zhang: I was pretty starry-eyed. I thought social media could give voice to the voiceless, expose corruption and help people coordinate action together.

Thomas Dinki: Today, after years of researching social media, Zhang has a much more clear-eyed view of its promise and its pitfalls. As an associate professor in the Department of Communication, she’s studied everything from how U.S. news media amplified Russian accounts during the 2016 election to how AI-generated content is impacting social media today. Her computational social science work examines how these technologies shape what people see, feel, think and do, and how that in turn influences politics and society.

Welcome to Driven to Discover, a University at Buffalo podcast that explores what inspires today's innovators. My name is Tom Dinki, and in this episode, I'll be speaking with Professor Zhang about why social media hasn't always lived up to its early promise, what forces are at play on our feeds today, and whether AI is about to shake everything up yet again.

Dr. Zhang, thank you for being here.

Yini Zhang: Thank you for having me.

Thomas Dinki: I think we all remember our first social media account. Mine was MySpace. What was your first exposure to social media, and how did you feel about it at the time?

Yini Zhang: My first social media account was on Renren, which was the Chinese equivalent of Facebook. And back then, I used that social media as just like, you know, pretty much every college student back then used it, mainly to keep up with friends and see what they were doing. And so it was pretty standard college behavior.

But at some point I realized that social media can be more than that. So I remember seeing like a protest in China that gained national attention via Weibo, the Chinese equivalent of Twitter. Basically, in this southern city in China, there was like a, you know, a proposed petrochemical plant, and locals vehemently opposed the plant. And they started to spread information online and they started to organize protests offline. And I realized that, wow, social media can be more than that, right? Social media can be more than just, you know, keeping up with friends, and then checking out what they are up to, and it can actually be used to do something that's much bigger than that.

And so that's, that's my first impression of social media, and that's what I was initially drawn to. You know, learn about more about that technology and its potential for like, you know, helping people work together, hold the powerful people accountable and organize collectively. And today, more than 20 years later, I would put 10 caveats after that, you know, statement, right?

Thomas Dinki: So you enroll in the communication PhD program at the University of Wisconsin-Madison to study social media, which involved conducting surveys. But you were skeptical of that methodology. Why?

Yini Zhang: I was. I was not a particularly serious survey taker, and you know when researchers asked me to fill out a questionnaire on issues that I knew little or cared little about, I could quickly give an answer and move on with my life. And I suspected that I was not alone. And then that got me asking what exactly the answers are, right? So it might not be genuine opinions that people might have. So that's why computational social science became so exciting to me because I was like, well, we don't have to ask people. We just need to go to online spaces and see what millions of people are actually talking about in their day-to-day lives in those online spaces, and get data and extract information from that data to learn about their opinions, their real opinions. But now, over, you know, after 12 years of research, I have grown more humble about this methodology. I would say every methodology has its own strengths and weaknesses, and every methodology is just a particular window into human behavior.

The 2016 election was kind of like a significant moment for me that furthered my interest in social media and politics. Back then, although the national polls came pretty close to the popular vote, several crucial state polls really were really off.

Donald Trump: And we are going to make our country great again.

Yini Zhang: Back then, the candidate Trump was so effective at using social media, using Twitter to dominate the news cycle, so that got me really interested in like the question of, what is the role of social media in politics, and how does, you know, kind of like online attention become amplified and reach the entire public.

Thomas Dinki: And in the wake of the 2016 election, you did some work on Russian social media accounts that eventually ended up in the Mueller report. Can you kind of tell us about that work?

Yini Zhang: Absolutely. That was one of the rare moments when you realize that well, like, you do research, and actually some people outside academia are paying attention.

[News broadcaster]: The first charges in the special counsel's investigation into Russian interference in the 2016 presidential election [inaudible].

Yini Zhang: So basically, after the 2016 election, it was found out that the Russian government sponsored an agency called the Internet Research Agency to operate an online influence campaign during the 2016 presidential election cycle. And later on in 2017, several social media platforms turned in their data to Congress, which was later released to the public. Our research team took the data, and we were curious about this one question: Did any Russian-sponsored content get into U.S. news coverage?

So initially, we looked at 33 major U.S. media outlets; those include USA Today, Fox News, The Washington Post and some partisan outlets like Huffington Post and The Daily Caller. And what we found was that in 32 out of the 33 outlets we searched for, Russian IRA-linked tweets showed up, and then later on we looked at a much more expansive list of outlets, and we found that 71 out of the 117 outlets we looked at actually embedded IRA-linked tweets.

So basically, those accounts infiltrated into the American public discourse, and here's the tricky part: They did not invent some spectacular, like, fake news. Instead, they just pretended to be Americans. They talked about issues like, you know, race, police brutality, immigration, and they try to be as divisive as possible. They try to, you know, drive a deeper wedge into the already, you know, very deep division in the American public, and that's how they exploited the existing fractures in the American public life.

Thomas Dinki: So if you were an American reading these news outlets, you might think that these accounts are your average Americans just voicing their opinion on American politics.

Yini Zhang: Absolutely. Yeah, journalists back then, I would say, did not have this great awareness of any potential influence campaign going on in the back end that can influence social media behavior, because people would go to, you know, would go online and then see, hey, this is, like, you know, like an ordinary American talking about the issue of race, talking about the issue of immigration, and that seems to be like, you know, an interesting take. Let's embed it in a news story to make it more, you know, interesting, to make it more attention-grabbing.

Thomas Dinki: It's fascinating how you're able to track how all these ideas spread across social media and then can influence news media. How do you actually go about doing the work of tracking these posts and finding out the ways they move? Maybe like within networks on social media.

Yini Zhang: So I would say the simplest way I would describe my methodology is this: I use computers to see patterns that I cannot see by reading, you know, posts one at a time. Typically, the first step in my research pipeline is to get data. Either I get data from social media companies through some like, you know, lengthy application processes, or you purchase data from third-party data brokers. So that's the first step, right? Getting data. The second step is, you know, for me or my team to sample a small data set from the huge volume of data that we have collected. I want to see with my human eyes, you know, what the data is telling me and what people are actually talking about before I let machine do the work. If I do not have a strong prior about what people are talking about, I can let computers find patterns for me, and then I will interpret the patterns.

And then after that, it will be typically network analysis because it's about, you know, who follows whom online, and that can tell us a lot about who people are, because people do mingle with the like-minded others online, and then find communities online to get a sense of, you know, what community is talking about a given idea, a given issue. So basically, if you think about the methods, they might sound like slightly intimidating, but the questions we ask are really, really simple, or like are really human, right? Like who's talking and who's talking about what, who's responding, and how ideas travel from one small corner in an online space into the entire, you know, public sphere to the entire media ecosystem.

Thomas Dinki: That's really fascinating and a good way to get into the subject of social media today. For example, it's estimated over half of all web traffic comes from bots, including AI-powered bots. How do you know that the posts you're analyzing were made by humans?

Yini Zhang: Well, I have to say that sometimes we just don't, and that is important to acknowledge. Right now, we have, as an academic community, developed some sophisticated tools to identify bots based on different indicators like, you know, the posting behavior, like, you know, the network structure, like who an account follows, and some other, you know, characteristics. One great example is the Botometer, which was developed by academics at Indiana University, but that tool can make mistakes, right? It can classify a human as a bot or a bot as a human. So there's this kind of methodological issue there.

I think on a deeper level, there is this conceptual issue that is, you know, what is a bot and what is a human being, and that line is becoming increasingly blurry. You think about like, you know, a human being using some automated system to schedule a posting automatically. Is that account a bot account or a human account? And you think about a human being posting something that is AI-generated, then is that account a bot or human, right? So there's this kind of like fuzzy conceptual kind of, like, boundaries that we have to think about here.

So as a researcher, what can we do? Well, we can think about, well, can we identify some coordinated account behavior, not necessarily bot behavior, but accounts that coordinate together, like Russian IRA accounts. They coordinated with each other to spread some, you know, narratives in order to manipulate the public. And can we identify some suspicious accounts and do some robustness analysis and see if we include those accounts, and if we don't include those accounts, can our results hold? So we can do all those things as researchers, as kind of like this imperfect solution to this really tricky question of dealing with bots.

Thomas Dinki: So at the advent of social media, it was thought that it was going to be a sort of town square where ideas could be debated and exchanged. Does your research find that productive discourse does still actually happen on social media?

Yini Zhang: Yes, social media is like a town square, but in this space, everybody has a microphone, and some people have tremendous speakers that can, you know, speak to many people at the same time, and there are billboards everywhere—you know, that's ads—and there are people who constantly yell at each other. So is that, like, a, a perfect town square? I don't know, right?

Also, we know that algorithms are quietly making decisions for us, right? They decide, you know, to a large extent what we see and who we hear from. So platforms have accumulated tremendous power these days, but we have seen through my research that social media is very good at bringing visibility to obscure issues or like issues that were obscure, such as, you know, sexual violence. And social media was effective in drawing attention to this issue of police brutality and, you know, racial injustice. So social media can do a lot of good things, but we need to be wary of the negative outcomes that can come from this same structure.

So the same structure that produces those desirable social outcomes can also reward conflict, can also encourage identity performance, reward attention-seeking. You know, encourage people to say provocative things online in order to like, you know, get engagement, get followers or get likes, right, and views. So social media might give everybody a voice, but it does not mean that social media fosters understanding and mutual respect.

Thomas Dinki: Is there any evidence that social media discourse actually changes public opinion or changes how we act offline?

Yini Zhang: We have to define the term change, right? What is change, and how computational social science help us understand change? I would say, in recent research that has not been published, we found that when there's greater discussion about the Second Amendment rights, there would be more gun sales. The pattern does not necessarily mean that, well, somebody read like a social media post about Second Amendment right and then jumped off the couch and bought a gun, because our data simply does not allow us to show causal relationship in that case.

So what the finding suggests is something happens before another thing, and that's a temporal relationship there. But how can we understand causal mechanisms? We have to rely on experiments. These days, I have started to conduct surveys and experiments to understand why people do things that they did. And so it's a little bit ironic because, you know, 20 years ago I was really skeptical of surveys and experiments, and now I need to rely on surveys and experiments to help me understand, you know, the questions that arise from computational social science research.

Thomas Dinki: They sort of fill the gap

Yini Zhang: Exactly. That's how, like, methods, they kind of, like, they complement each other.

Thomas Dinki: Now, of course, AI is changing social media, which you have been researching. What are you looking into exactly when it comes to AI and its impact on social media, and what are you finding so far?

Yini Zhang: AI has become this exciting new research topic for me because I see AI not just as a tool. I see AI as this emerging media and communication technology that can tremendously shape public life and communication in the next two decades maybe, in the same way that social media did the same thing to us in the past two decades.

One particular question that we have started to look at is to what extent people can accurately identify the amount of AI content in their information environment. So to do that, we conducted an experiment where we controlled for the amount of AI-generated images in their social media feeds. So, for example, one group of people might get zero AI-generated visuals, just pure news images, and another group of people will get like, you know, like 10% and or like, you know, 40% and 80%. So that's a study that is currently done in collaboration with Dr. Jiyoun Suk at the University of Connecticut, and what we found is that people overestimate the presence of AI-generated content in their social media feed, and even when there's zero AI in their feed, when all the images, all the news images, are actually from mainstream news outlets, they still think there's about, you know, 50% of AI-generated content in their feed. To make it worse, when the content is against their pre-existing belief, they would even further overestimate the proportion of AI-generated visuals in the feed. So you know, if the content is not agreeable, if I don't like the content, then it must be AI-generated, or there must be a lot of AI-generated content there.

So what the study tells me is that AI is not just like an information problem. AI really contributes to distrust in the social media environment because we ask people, well, like, to what extent do you trust this platform? And when they perceive that there's a lot of AI, they are less likely to trust the platform that they, you know, get content from. I think that speaks to a deeper epistemic problem there, because these days when you see content online, and either text or visual content or video content, you don't know who generated the content, you don't know who posted the content, you don't know whether the content is human or AI-generated. So there's this problem of trust, general trust. Who do we trust? What to trust? And that I think is a deeper epistemic problem that we are struggling with, and we will be continuing to wrestle with in the next, you know, many years to come.

Thomas Dinki: So you started out being very optimistic about social media and what it could do for democracy. Are you totally pessimistic now, or do you still have some optimism?

Yini Zhang: So there are both, you know, promises and pitfalls associated with any technology. I think the problem is, you know, how do we steer this technology and how do we guard against the, you know, the negative consequences? So, and then based on my research, I feel like, you know, again like social media allows people to, you know, talk to other people, social media allows people to follow each other. But again, like, exposure is not understanding, and being able to follow each other is not the same as building real connection. And those things do not automatically produce a better society. Those things do not automatically enhance democracy.

And there are a lot of things that we need to watch out for, like, you know, harassment on an individual level, or like, you know, cybercrime, cybersecurity, misinformation, disinformation, all those issues that we have identified in the past 10 years or so are here to stay, and these days, you know, AI is making those problems, you know, more intractable to deal with and to solve. So I guess I have become like professionally incapable of being either techno-optimistic or techno-pessimistic. I guess I am right in the middle. I guess I'm more like techno-curious and with a very, you know, cautious approach to it. And I think that's, you know, a nice place for a researcher to be and maybe, you know, for anybody to be in this day and age.

Thomas Dinki: Dr. Zhang, thank you so much for joining us.

Yini Zhang: Thank you so much. Not everybody gets a chance to talk about their work at length, let alone on a podcast. It's a pleasure.