Joshua A. Tucker is Professor of Politics and Professor of Data Science at New York University. He is the Director of NYU’s Jordan Center for Advanced Study of Russia and a Co-Director of NYU’s Center for Social Media and Politics. He spoke with SSRN about the relationship between social media and politics, and how different types of data help us understand the connection between online and offline behavior.

Q: In the past decade plus, you focused on studying the relationship between social media and politics, as well as ways to use social media data to study politics. How did your career lead into this area of focus?

A: For the decade and a half before, I was working on mass political behavior in post-communist countries, and I had done a lot with elections, voting and partisanship, and public opinion formation. In the 2000s, there were a series of big protests that took place after serious instances of electoral fraud in Eastern Europe and the former Soviet Union, and I got interested in those. One thing I heard frequently from scholars working in this area was that there was never going to be one of these protests in Russia. Then one day in 2011, we woke up and there were over a hundred thousand people on the streets of Moscow protesting against electoral fraud. I asked friends of mine in Russia what was going on, and I heard repeatedly about Facebook and how these protests were being planned online. That was the first time that I started to get interested in the idea from a substantive standpoint, that maybe social media was something that we were going to need to think about as political scientists.

Concurrently with that, I had this absolutely brilliant graduate student, Pablo Barberá, who wrote a paper showing that you could use network data to predict people’s ideological orientation or partisanship. He was proposing [that] you could do this with Twitter data, trying to estimate people’s ideology not from what they were saying in their tweets, but rather based on whom they were following. Now, in 2024, from the vantage point of large language models, you’d think, ‘Well, of course, we could just use the content of their tweets to estimate ideology.’ But at that point in time, the field of text-as-data was much less developed, especially in political science. As a result of Barberá’s paper, though, I for the first time started to think seriously about the possibility that we could use social media data to actually study politics.

The third thing that happened around the same time was that there was a call for proposals from the National Science Foundation (NSF) to support “outside the box” interdisciplinary work. We put together a team of faculty at NYU and put in a grant proposal. When you write NSF proposals, you always think you have a very low chance of getting them because of how competitive the process is, but we actually got this one! We began to embark on this experiment, not just studying new questions – how social media impacted political behavior – and using new data, but also an experiment in trying to bring the lab-based approach to research that is so common in the natural sciences into the study of politics. That’s what we’ve been doing for the last dozen years now.

Q: As faculty co-director of NYU’s Center for Social Media and Politics (CSMaP), and as a researcher, what data do you focus on collecting and analyzing, specifically during an election year like this one?

A: We tend to study the relationship between social media and politics in a couple of different ways. The big challenge here is that social media is huge. There’s a ton of social media data, and especially because it is optimized for search, it’s easy to find examples of anything. However, moving beyond anecdotal examples to understand what’s happening at scale is really challenging. Also challenging in this context is trying to rigorously test causal relations about the impact of social media on political outcomes.

These are complicated tasks. We’ve ended up using two different sources of data to try to address these tasks. One source of data is the social media data itself… and that has all sorts of challenges around it. There are legal challenges around it. There are ethical challenges around it. There is the logistical challenge of collecting this data at scale. The other big bucket of data is survey data: surveying people about their relationships with social media and politics. Then we’ll layer experiments into both of these things.

For example, my lab has done a bunch of deactivation experiments, where in order to get causal relationships, you take people who are using a social media platform and get them to enroll in a study. You then randomly assign one set of participants, the treatment group, to stop using the social media platform or to reduce usage, while the other set, the control group, continues to use the platform. Then we’ll use survey data to look at differences and changes over the period of time.

We’ve also been running a survey for many years where we have a panel of people who we’re surveying repeatedly, but also over the years have had the opportunity to link their surveys with their social media data, which they consent to provide to us for research purposes as well. A lot of our most important papers were written because we were able to link up survey data, which allowed us to know things about people, know their opinions, know their preferences, [and] also to know their demographic characteristics, with actual social media data.

That’s sort of the universe of what we can do. There are the observational studies where you’re trying to figure out what’s happening on the platforms. There are more traditional social scientific studies, where you’re running experiments and using surveys to get at changes in people’s attitudes. Then we’re trying to marry those two and combine them.

Q: How do you use certain data and information to try to understand the interaction of what happens online and offline?

A: There are a couple different ways you can do it. The simplest way is starting with surveys, where you ask people about their social media usage. And it’s very similar to traditional social science research that we would do. The other thing you can do is bring digital trace data into these kinds of surveys… saying ‘Okay, well, if we change the type of media that people are seeing online, what does that mean about their attitudes about things that are happening offline’?

We ran this deactivation study, the idea of another brilliant graduate student in our lab, Nejla Asimovič, where we had people stop using Facebook in Bosnia and Herzegovina. This was at a time of commemoration of genocide that had taken place in the Yugoslav War. We thought, based on the literature, that people who were off of Facebook would have lower levels of ethnic polarization [and] would express lower levels of hostility towards other ethnic groups. We thought this because, around the time of this genocide commemoration, we knew there would be a lot of negative things that people were being exposed to online.

We actually found the opposite, which was really surprising to us. The people who stayed on Facebook had lower levels of antagonism towards other ethnic groups in Bosnia. In this study, we asked the people who were in the deactivation study, “what did they do with the extra time?” Some people said they went on Instagram — a substitution of other social media. But tied with Instagram for the most popular thing people said was that they spent more time with friends and family. That’s an offline behavior.

What we thought maybe was happening — because Bosnia has this terrible history of ethnic cleansing — is that the people who were online were still having at least some contact with people from other ethnic groups, whereas maybe the people who were offline at the time of this genocide commemoration were perhaps only talking with other people from their own ethnic group. We thought ‘if this is correct, then this effect – having higher levels of ethnic polarization if you are off of Facebook than if you stayed on it – should probably be driven by the more ethnically homogenous parts of the country.’ We reran our analysis, but this time separately for those in the more and less ethnically homogenous parts of the country, and it turns out that this was exactly right. That’s a great example of how the offline and online environments can have these mutually interactive effects.

One of the great challenges of working in the information environment is that it’s hard enough to study in one country what’s happening on one platform. Of course, people don’t live on one platform: they’re watching TV, listening to podcasts and talk radio, and they’re also talking to their friends. This online-offline tension is an important one, but it’s also a really fascinating subject for research moving forward.

Q: So far, we’ve been talking about big picture ideas relating to social media and politics, but I’d like to zoom in a little bit. What are some specific studies or research projects that you’ve been conducting recently that you’re particularly excited about?

A: There’s a lot of them, because we have a very active research center here. One is that we have been interested in people’s ability to identify the veracity of news, and trying to figure out what sorts of interventions make people more or less likely to be able to correctly ascertain the veracity of news articles.

We set up this pipeline where we put together five streams of media sources: left leaning mainstream media sources, right leaning mainstream media sources, left leaning low quality news sources, right leaning low quality news sources, and then low-quality news sources that we couldn’t tell if they were right leaning or left leaning. Every morning, we would take the most popular article from each of those five news streams, and we’d send them out to 90 different people, as well as to professional fact checkers. We would take the mode of our professional fact checkers to be the “correct” answer (i.e., that the central claim of an article was either true, false or misleading, or it was impossible to tell), and then we could look at the variation in how different people were able to match the professional fact checkers or not, in recognizing if the central claim of the article was true (or false) or not. Then we were able to look at the impact of a whole host of different interventions in people’s ability to correctly identify the veracity of these articles.

One of the things you see on all these digital literacy interventions is, “if you’re not sure about it, go online and look up and try to get more information… from a reputable source.” One experiment that we ran was to have a treatment group search online for more information about the article and then assess the veracity of it, as opposed to the control group that just read the article and then assessed its veracity. We ended up running five studies to do this, [and in] study after study after study, when you went and searched online, you were more likely to believe that a false news story was true, not more likely to correctly identify a false news story as false.

The big takeaway from that is how important it is to rigorously test these digital literacy interventions. I think it is probably a good idea to tell people to search for more information from reputable sources, but it’s that ‘reputable sources’ part of it that’s super important.

What we’re doing now is trying to extend this research. We’re running a brand new study where we’re going to have people who go to traditional search, but then we’re going to send another group of people to generative AI, and we’re going to see if generative AI is better than the traditional search or if it’s even worse.

Q: Your most downloaded paper on SSRN is “Social Media, Political Polarization, and Political Disinformation: A Review of the Scientific Literature” which has been cited, shared by news sources, and downloaded consistently since its posting on SSRN in 2018. The core topics you visit in this paper remain extremely relevant today. Looking back two presidential elections and a global pandemic later, are there any new problems that you couldn’t have predicted? Or are we facing similar problems now, just different variations of the same issue?

A: In the conclusion of that paper, we had a couple of recommendations. We said the literature was too focused on Twitter. [Since then], people have started to diversify and look more at different types of platforms. At CSMaP, for example, we’ve done research on Facebook, and we’ve done research involving Reddit. We have a whole set of papers involving YouTube, and we’ve just written our first paper about NextDoor. Since 2017, [researchers] have gotten better about not just studying Twitter because it’s the easiest thing to study.

The other big recommendation was that literature was overwhelmingly studying the United States, and beyond the United States, it was overwhelmingly about other advanced industrialized democracies. I think that probably still characterizes a lot of literature, but there’s progress being made in that regard. In our lab at CSMaP, we’ve been trying to do this: for example, the Facebook deactivation that I talked about in Bosnia, we have now replicated that in Cyprus. We did a WhatsApp deactivation now in Brazil, and now are extending that work in South Africa and in India. There is a good recognition that we need to diversify out beyond the United States.

At the time we were working on that SSRN paper, people were coming off of this original euphoria that ‘social media was going to help spread democracy all over the world,’ from when I originally got interested in it. Then, the pendulum swung way back, and people were terrified about social media’s threats to democracy. Jumping forward seven years now, in terms of the kind of threats that we’re concerned about, they’re still fairly similar. We’re worried about whether or not people find themselves in small, isolated communities online, where it makes it easier to radicalize people towards extremist views, whether people find going online a hostile experience, whether or not the experience of social media makes people dislike people from other political parties more, and whether in authoritarian countries, regimes can use it to control the information environment of their own people.

The part that people may not have anticipated in 2017 was questions about the effect of the rise of generative AI, artificial intelligence and large language models — how that would shift people’s attention, but also how the large language models would interact with social media. Social media lowered the cost of sharing content dramatically. Prior to social media, to spread content far, you needed to be on a television show or have access to a newspaper or write an op-ed. With social media, suddenly anyone could share content. But you still had to produce the content. Generative AI has lowered the cost of producing the content. These are two parallel processes, but they are also ones that I think can interact with each other. We’re only at the beginning of understanding how that’s going to go.

Q: How do you think SSRN fits into the broader research and scholarship landscape?

A: I’ll let you know what it’s been for us. There are disciplinary versions [of preprint servers] and political science doesn’t quite have one. So SSRN has been one that a lot of political scientists use, and we use often in our lab. We’ve definitely learned that it is much more effective than just putting the paper on your website — to have it in an archive where people can discover it.

A second thing goes back to the article we were just talking about. This review of the literature was never intended to be an academic article. It was a report commissioned by the Hewlett Foundation, and the Hewlett Foundation put it on its website. I thought people in the policy community were going to see it on the Hewlett website, but I also wanted people to see it in the academic community. I thought that maybe we’d get a few citations out of it, and [decided] to throw it up on SSRN, on a whim. And now it’s been downloaded over 40,000 times and continues to be cited all the time. In that sense, it filled this really nice niche: we had something that we didn’t write to be an academic publication [and] weren’t going to send to journals. It’s a nice home for papers like this that don’t have a natural fit in academic journals but can still be useful to researchers and policy makers.

The third use: we’ve had a couple of papers that we have had trouble getting published, but we’ve put them up on SSRN and they’ve gotten read and cited. I think that’s been the other real benefit of SSRN for us: papers that for whatever reason have had a tough time landing in journals have been able to have a life on SSRN as we slog through the publication process.

More About Joshua Tucker

Joshua A. Tucker is Professor of Politics, affiliated Professor of Russian and Slavic Studies, and affiliated Professor of Data Science at New York University. He is the Director of NYU’s Jordan Center for Advanced Study of Russia, a Co-Director of NYU’s Center for Social Media and Politics, and was a co-author/editor of the award-winning politics and policy blog The Monkey Cage at The Washington Post for over a decade. He serves on the advisory board of the American National Election Study, the Comparative Study of Electoral Systems, and numerous academic journals, and was the co-founder and co-editor of the Journal of Experimental Political Science. His original research was on mass political behavior in post-communist countries, including voting and elections, partisanship, public opinion formation, and protest participation. In the past dozen years his work has increasingly focused on the intersection of the digital information environment, social media and politics.

You can see more work by Joshua A. Tucker on his SSRN Author page here.

Sara Gerke is an Associate Professor of Law at the University of Illinois College of Law. Her research focuses on the ethical and legal challenges of artificial intelligence and big data for health care and health law in the United States and Europe. She spoke with SSRN about the legal and ethical challenges of integrating artificial intelligence into the medical field and how much farther we still have to go.

Q: You’ve led and been part of research projects that work on things like artificial intelligence (AI) in healthcare and the legal and ethical implications of integrating new and cutting-edge technology into medical practice. What has driven your interest and motivation within this field of study?

A: I moved to the U.S. about six years ago. Previously, I was the General Manager of the Institute for German, European and International Medical Law, Public Health Law and Bioethics of the Universities of Heidelberg and Mannheim, and I was also doing a lot of grant writing. I already felt like we were getting to the point where we have to submit applications for technology, because that’s the new world, where it’s heading towards. It takes quite some time until you actually can execute a grant. I somehow got to know about this project [in the U.S.], and I was really excited about it. It would have taken probably another two years until I could carry out something similar in Germany. Long term, I always thought to pursue an international career so that I can write more articles in English and really spread my work around the world. I applied for that position, was very lucky to get it, and was responsible for the day-to-day work of the Project on Precision Medicine, Artificial Intelligence, and the Law (PMAIL) at the Petrie-Flom Center for Health Law Policy, Biotechnology, and Bioethics at Harvard Law School.

I feel these days everyone is very aware of AI, at the latest when ChatGPT came on the market. But six years ago, we were one of the very small groups who really looked at AI and digital technology at that time in the healthcare field from an ethical but also legal perspective.

Q: Your most downloaded paper on SSRN, “Ethical and Legal Challenges of Artificial Intelligence-Driven Healthcare,” written in 2020, highlights several main concerns regarding AI in healthcare, both in the legal and ethical sense. In the four years since you wrote this, which legal or ethical problems do you think have made the most progress?

A: I don’t think we made enough progress in those years since I wrote that paper. This paper has been cited so much because we were essentially the first who wrote a paper on this topic. At that point, it didn’t have much interest yet. But what we are seeing, comparatively – because I do a lot of comparative law between the EU and the U.S. on the regulatory front – is a lot has happened since then. The Food and Drug Administration (FDA) has thought a lot about how to regulate AI and machine learning (ML)-based medical devices. Also, we have seen in the EU just recently the AI Act, the worldwide first regulatory framework for AI.

In the last probably two years now, people are very aware of questions of liability and liability risk when introducing those tools into the healthcare system and into the hospitals. The FDA has already authorized over 950 AI/ML-based medical devices. And frankly, not all of the AI tools are considered or classified as medical devices. We have a lot of in-house developments and other particular clinical decision tools that fall outside of FDA regulation. Those tools [have] already been implemented in healthcare. In recent years, these questions of ‘who’s going to be held liable if something goes wrong?’ are much more pressing.

Q: How did working on this research early on inform what you do now?

A: I feel very fortunate that I have been working in this field for such a long time now to build my expertise. I can now adapt to new challenges relatively easily because I have a long experience of going deep into the legal and ethical issues of AI. I’m doing a lot of interdisciplinary work. I work with physicians, with engineers, and so I can always get new ideas from those collaborations. Every day you read something [new] about AI: it’s hard to keep up at that pace. So, some groundwork of knowledge is really helpful to be able to catch up on all the new developments happening every single day.

Q: You’ve done a lot of work regarding medical AI applications used for self-diagnosing and the importance of labeling these apps properly. Talk a little bit about why it’s important to educate people about the function and purpose of self-diagnosing AI apps like these.

A: I’m still doing a lot of work on the questions of labeling, because I think it’s so important. Labeling is just one small piece of this entire puzzle. I am a believer in labeling AI, in particular in the healthcare field. Imagine two different types of tools. You can first imagine a tool which is being used by a physician in a hospital. The label is directed to the healthcare provider and [those] who need information about the tool to assess whether it makes sense to use that tool safely on their patients.

On the other hand, we have a lot of direct-to-consumer AI tools, including apps. Here, they also need to get some type of information. This information might potentially be a little different in detail. A consumer might care about other things – like privacy, if this is safe to use, or if data is properly protected – while a physician might need to get more granular information about if this dataset [has] been properly tested and breakdowns of race, ethnicity, location, age, gender, etc. That information is really important for physicians to know, and unfortunately, we don’t have proper standards in place at this point which require manufacturers to disclose that.

The Office of the National Coordinator for Health IT (ONC) has recently put out a rule and requires some type of transparency and some kind of nutrition label. [It’s] really nice to see that regulators are now finally doing something in regard to labeling, but it’s just a tiny amount of tools that are being covered by that rule. I’m still waiting and hoping for the FDA to change that. At some point, hopefully all AI tools have some type of label, so that users are getting proper information. Sometimes one gets the criticism, “well, no one reads labels,” but it’s better to have that information as an option to read it, if you want to. I think there is a necessity to require that disclosure.

Q: You’ve pointed out in your work that while many of these apps might be labeled as “information only” tools rather than actual diagnoses, people sometimes misperceive them as real diagnoses. Do you see this misperception as an example or indicator of some sort of larger issue with how medical AI tools are understood by consumers?

A: The issue that we see in some of these apps is that they – for example, the electrocardiogram app from Apple – are actually information tools and not [meant] to make a diagnosis. But if people are using those tools… for [something] like skin cancer screenings, they usually tend to believe whatever the AI tool gives them. [They] might just skip a doctor visit because they are scanning with their app, and [think], “oh, it seems I don’t have any cancer. Everything looks good, so I don’t need to necessarily go to the doctor right now.” We see a lot of empirical data on that, that the consumers are perceiving those tools as diagnostic tools. But if one looks at the language, all of those manufacturers are clearly articulating that a direct-to-consumer tool, often available without any prescription, doesn’t replace going to the doctor.

I think we need to have measures [and] more education around it. The majority of health AI apps or general health apps are not being reviewed by the FDA: they fall outside of the Federal Food, Drug, and Cosmetic Act. So, it is hard for consumers to assess whether this app is reliable or not, because consumers usually just put into the App Store what they are looking [for], and whatever pops up first, they likely are going to download.

Q: AI has developed very quickly in the past few years, and it has raised problems that nobody really had to think about before. Because of this, there’s a retroactive response to certain problems. How do you think researchers can begin to anticipate where the field is headed, to ensure that AI is used responsibly in healthcare?

A: I personally don’t have a crystal ball, and I think that’s the problem. Because honestly, if you would have asked people maybe three [or] four years ago, I don’t think they would have suggested that we’d come [this] far so quickly. We have these generative AI tools that have incredible capacities but also raise a lot of new issues which we have not anticipated. Once you get it out of the box, it’s hard to get it back into the box.

That’s a problem now, because retroactively making laws around this is really challenging, and we are not seeing right now in the U.S. that that’s going to necessarily change. I think the approach in the U.S. is going to be more like a mosaic style. We have different regulators, and everyone is going to do some stuff in their wheelhouse. Hopefully, there will be enough collaboration and understanding that it’s going to be a mosaic or puzzle to be completed, rather than overlapping and making it much more complicated for stakeholders to understand and oversee.

Q: Are there any papers, projects, or research you are working on right now that you’re particularly excited about?

A: I always have a lot of projects going on, because AI keeps me busy. I have several research projects I’m involved in. I’m leading two projects in the ethics and legal field, which are called CLASSICA and OperA, and they are funded by the European Union. I’ve also been one of the PIs (principal investigators) of an National Institute of Biomedical Imaging and Bioengineering (NIBIB) and the National Institutes of Health Office of the Director (NIH OD) grant on technology. There are – in particular in the two projects funded by the European Union – clinical trials being carried out. One is for colon cancer prevention. AI tools are being tested to see if they are going to be helpful in the long run. Similarly, [in] the CLASSICA project, that’s a project on AI-assisted surgery. The surgeons are testing an AI tool that can predict whether tissue is benign or malignant in real-time during a surgery.

I’m not involved in the clinical trials, but for me, it’s really exciting, because my team is looking at some of the legal and ethical issues, such as, “do surgeons have any reluctancy to implement such a tool and use it in the operation room?” [and] “are there any liability risks that they are worried about?”

My work keeps me busy in the liability space, but also in regulation and the questions of, “how should regulators in the U.S. and Europe regulate more complex tools like generative AI?” That’s going to be a real challenge.

There are many questions, and in the liability space in particular, the more sophisticated the AI tool becomes, the more interesting and unsettled the question is. “How does tort law deal with an AI tool that practices medicine?” An AI tool is not considered a legal person at this point. So at some point in the future, you might have an AI tool that is so sophisticated [that] it’s a standard of care and is implemented in a hospital. It might be that you can’t find any human fault in the physician using the AI tool: the physician needs to rely on the output of the AI tool because it’s so complex. And then there’s this question of, “if harm occurs, who’s going to be held liable if you can’t find human fault in the physician?” Because it was totally fine of the physician to rely on the AI tool in the first place. These are questions which we need to tackle in the future, once such AI is going to be implemented in the healthcare field.

Q: How long do some of these clinical trials take?

A: The ones carried out in the projects I am involved in last for several years. These are long-term projects spanning four to five years, and I think they are [some] of the rare occasions where clinical trials are actually being carried out in the AI field. The majority of AI tools, in fact, especially in the U.S., have not undergone any clinical trial studies. How the Federal Food, Drug, and Cosmetic Act works is that there is a pathway, which is called a 510(k) pathway. What you need to show as a manufacturer is that your device is substantially equivalent to another legally marketed device, which usually does not require any clinical evidence. We have seen that the majority of the AI-based medical devices that were authorized by the FDA went through the 510(k), so in most cases, there is no necessity to show any clinical data. And so that’s what we are seeing.

Clinical trials are rare in the field. But, of course, clinical trials should not hamper innovation. It’s hard, because if you have a so-called adaptive AI tool that continuously learns, you could have a clinical trial, but how much of the results of the clinical trial data will still help you in the long run, [given that] the tool is constantly changing and adapting? If we are getting fully adaptive systems… it’s going to be even more difficult to make sure that they stay safe and effective. You will probably need to have an ongoing monitoring system in place to be able to tackle that issue.

Q: There are many people that are still wary of the idea of AI being used in medical practice. It feels new, and people can be nervous about things that they haven’t experienced before. What would you say to people with those concerns?

A: In general, AI tools can have a lot of potential. I think one also needs to really see what type of AI tool it is. If a physician is using it in practice, I think it’s really important to be very frank with the patient about it. What are the benefits? What are the risks? What may be issues which are unknown, so that the patient could have the choice to decide whether they want it to be used in their care, at least for the transition phase of where we are. Because at this point, the use of AI is not yet the standard of care. But, of course, the standard of care evolves. During this transition phase, it’s going to be essential that physicians communicate properly to patients.

Q: What do you think SSRN brings to the world of research and scholarship?

A: I think SSRN is great because, first of all, it’s a known platform. It’s for free. Everyone can use it: it’s open access, so that’s great. Also, we can upload forthcoming paper drafts early on, so that this can be spread across disciplines to other scholars before it’s even been published. I think these are all great advantages for scholars in general and give people access to the work as soon as possible.

More About Sara Gerke

Sara Gerke is an Associate Professor of Law and Richard W. & Marie L. Corman Scholar at the University of Illinois College of Law. Her research focuses on the ethical and legal challenges of artificial intelligence and big data for health care and health law in the United States and Europe. She also researches comparative law and ethics of other issues at the cutting edge of medical developments, biological products, reproductive medicine, and digital health more generally. Professor Gerke has over 60 publications in health law and bioethics, and her work has appeared in leading law, medical, scientific, and bioethics journals. She is leading several research projects, including CLASSICA (Validating AI in Classifying Cancer in Real-Time Surgery) and OperA (Optimizing Colorectal Cancer Prevention Through Personalized Treatment With Artificial Intelligence), both funded by the European Union.

You can see more work by Sara Gerke on her SSRN Author page here.