 Hello everybody and my name is Jan Piasecki and I would like to officially commence the Twitter research ethics online seminar on the web immunization project. I can see that we have participants from all around the world that's really great. Thank you for joining us and before before I introduced our guests and our speakers I would like to remind you about two things. So the first thing is that our research project is funded by the EEA grants and operated by the National Science Center in Poland and moreover our session is being recorded and the video will be later on available on our website and on our YouTube channel and right now it is my supreme pleasure to introduce to you our guest and moderator of the seminar Dr. Elizabeth Buchanan and she's a director of the Office of Research Support Services and Senior Research Scientist at the Marshfield Clinic Research Institute. For over 20 years Elizabeth's scholarship has focused on research ethics, compliance and regulations specifically about internet, social media and big data research. Our second guest is Dr. Nikolas Profares who is an assistant professor at Arizona State University School of Social and Behavioral Sciences. His research interests include users' understanding of social technical systems such as social media, social discourse about technology and issues of power and ethics in the digital landscape. And finally I would like to welcome not a guest but a speaker, the member of our Web Immunization Research Team, Professor Mikołaj Morzy, who is an associate professor at the Faculty of Computing at Poznań University of Technology. His research interests focus on machine learning and its application in natural language processing, complex network system and social networks. You can find more about our speakers by visiting our website. Also you can learn a little bit more about our research projects. Project and now I'm disappearing and I'm passing the microphone to Elizabeth Buchanan. Thank you. Okay. Can you see my slides? Okay. Yes. Okay. We can see you in the presenter view. So we can see also the next slide. Okay. Yeah, it works. Okay, very good. Okay. Well, hi, everyone. It's great to be here today. Many thanks to Dr. Piasecki and his team. Many thanks to Agnieszka for organizing this session and a big congratulations to the team on your successfully funded project. It's a great privilege to be here too with Drs. Morzy and Proferas. I will be providing some brief remarks and then moderating the session. I'm going to ask that you hold your questions. You can send them in throughout the session and then we'll address questions at the end. I just wanted to just say that I have no conflicts of interest related to this activity and any opinions or comments or recommendations that I expressed today are mine and do not reflect my employer, Marshfield Clinic Research Institute, Marshfield Clinic Health System. And like many people, you may have some unofficial disclosures. Zelda and Wally are very quiet so far today. They're peaceful and it's a fairly rainy, cloudy day here in Wisconsin, which is the little red area that you see there. And so typically though, as you know, we've all been working from home now for many, many months. And as soon as I start a presentation, of course, the mail or the post shows up and the dogs go crazy. So that's just my warning in case we have a visit from Wally and Zelda. So as Jan said, I've been working in this space for over 20 years. And I think the the research that's taking place in what we call the internet in its broadest terms, I think we're really seeing just growing complexity. A number of years ago, my colleague, Michael Zimmer and I were approached to do an entry for the Stanford Encyclopedia of Philosophy. And so we just finished our third substantive revision just this year, 2021. But what I was thinking about when when planning for this session is it's going back to the the initial definition that we came up with back in 2011-2012 when we first put the this entry together. And I think it still holds true. Conceptually and historically internet research ethics is most related to computer and information ethics. It includes such ethical issues as participant knowledge and consent, data privacy, security, anonymity, confidentiality, integrity of data, IP issues, community, disciplinary and professional standards or norms. And since this since early work, again, looking back into the the the 1990s even, some questions have emerged. And I kind of jokingly said so many questions on this slide. But these are just some of the questions that that we've identified over the years, some early on in our first iterations of IRE research. But as we glance at this list, it really is growing in complexity. Jan mentioned that when we were talking earlier about the new GDPR, and you know, researchers are now facing so many different complexities as we work in these different spaces. And I would say that the early questions, those that had to do with issues of what is what is public, what is private, those questions have certainly not gone away. And I would say they certainly haven't been resolved to any great degree of confidence. And I would say to the contrary, many of those issues that we identified way back when have have just grown either in intensity or in complexity. And I think it's fair to say that as we think about social media, we think about big data, machine learning, AI, those those have have stirred the pot of questions. And I think now we're at a point where we're seeing such complex considerations in internet based, internet, broadly speaking, research that that we're pushing very, very heavily back on epistemic and normative constraints. And I think as we as our collective interdependency on social media has grown over the years, our relationship with trust, truth, representation has also dramatically changed. And I think it's funny, we go back all the time to this silly cartoon from the 1990s about anonymity, right? Anonymity on the internet. Nobody knows you're a dog. But think about where we are now. And we're in this space of what we've moved from that notion of anonymity to precise levels of identification to prospective modeling of behaviors of prediction of prediction. And and it almost seems seamless, right? That these transitions have happened, you know, from from that space of anonymity to where we are now. And yet the ethics of the the transition, they're really the ethics are enormous. Excuse me. And a few years ago, I wrote a short response in public library of science. And we talked about, I was responding to an article that used a form of network, social network analysis called iterative vertex clustering and classification and IVCC. And the article was talking about using that that model to identify specific populations in large data sets. And I went back and forth with the authors. And, you know, it was a really fruitful conversation where we talked about methods and ethics. And how do we front load ethics in these kinds of research methods? It also got me thinking about the the true methodological power that we as researchers have at our disposal today. And going forward, I think that that power will increase. So as researchers working in on in between spaces on on our internets today, we have we have unbelievable access to data. But I also want to remind us remind us of the risks that that we also take as as researchers in these spaces. And, and I think, you know, we only need to think over the past, you know, five to 10 years of some of the the the famous or infamous cases of studies gone wrong to think about the the risks that possibly could occur. So as as I read more about Jan and his team and this project, and as they engage on this very ambitious project, they are going to be facing a tremendous ethical, regulatory, disciplinary challenges. And and so, you know, what what does a research team do? Where do you begin when you talk about front loading ethics? Well, in from a Western perspective, right, we would say, well, we start at the Belmont Report, we think about how we treat our participants in research, we think about risks and harms, and we think about the the benefits of research and also the burdens of research. But I also want to suggest that we think about an old an old document, another old document on being a scientist that really talks about the values that that are essential to our work as scientists and that we honor the trust that colleagues place in us, that we honor an obligation to do the best work possible and to embrace productive and honest work. And then this last one to uphold an obligation to act in ways that serve the public. And I think it's it's really in that last point where I was thinking about the web immunization project as really critical. And recently I was I was talking with a library system about distinctions of of myths and disinformation. And so as this team is starting this this project looking at this at misinformation, you know, what what are they going to uncover? And what I what I hope what I really hope from this team and I have great expectations for this project is I'm hopeful that that empirical research that empirical research looking at these forms of rhetoric looking at empirical data around misinformation disinformation. I'm really hopeful that that's going to contribute to a healthier discourse. I hope our our citizenry is is better informed as a result of the work that's going to be done. Excuse me. Okay, so just a few more comments. I think we do know a lot more about the ethics of conducting internet research or research on Twitter than we did just just a few years ago. I think we've really come a long way, particularly from some of the work that Nick Propharis has done. I think we're really starting to understand those those those very complex tensions that exist in those spaces of public and private. And what do participants expect? There there is a substantial literature across disciplines, but a trend that I've I've seen and I've heard in many conference presentations. And I think it's a little confounding for ethicists and for for some researchers that there's this presumed public nature of Twitter that's kind of become this default position. And I think we still need to tease that out. Regulatorily, I think that that probably is the correct framing of Twitter data. However, what about the ethics piece, right? I'm also concerned about the the ways in which that public nature enables us to, you know, within, you know, an hour or two, you know, get to develop an API, operate, you know, get this API to to to grab all sorts of data for us, right? It can enable we can have all sorts of research projects, you know, acting very quickly. And so I said this a number of years ago about these spaces that the shifting research landscape is complex, that the fact that data are coming from these myriad of sources, some of them intentional and some of them unintentional. I'm concerned too around about the concept of research bystanders or collateral subjects as sometimes called in these streams of data that that gets scooped up. One, once one's connections in a social media landscape do matter, right, they do matter, even if those those connections are distant or impersonal. Human subjects research broadly understood is fun, it is or should be fundamentally different in the age of data science. However, from a US regulatory perspective, the regulations really haven't kept up, they still haven't kept up with with the the age of data science, despite the rule revision in 2018. Methods such as IBCC rely on continuous data streams and and continuous analytics, and many of these data mining and analytics studies would be considered secondary analyses. The degree to which a researcher has access to identifiable data or the ability to ascertain information about individuals through, for example, re-identification techniques are used then as as determinants, right, of the level of risk and benefit in the current US regulatory model. So I just want to wrap up with a few words from a regulatory perspective. In the US, we did have a rule change in 2018. And I think a lot of researchers were hopeful that the revised common rule would address some of the technological changes taking place to some little little degree they did. I think most of us didn't quite see the the changes that we expected. But when we talk about these kinds of large data sets, we kind of have two different paths here from a US regulatory perspective. We we either fall completely outside of the common rule definition of human subjects research, and that the our definition is a living individual about whom an investigator, whether professional or student conducting research, obtains information through intervention or interaction with an individual and uses studies or analyzes the information. I'm leaving out the biospecimen piece for now. Or obtains, uses studies, analyzes or generates identifiable private information. So so the path would be, okay, does this does this data set, for example, comprise human subjects research? If not, there's no there's no ethical oversight, right? It's outside of the purview of the ethics committee or the IRB. If we agree that it does fit that definition of human subjects research, the next step would be, okay, what kind of research is it? Typically, the types of research we're talking about today would qualify for what's called an exempt determination status. And and once that determination status is is is exempt, that means IRB oversight effectively ends. So again, we're back to a point of there's no IRB or ethics committee oversight. And are we comfortable with that from a research perspective? Typically, what we would see, we would see that research in this vein would fall into an exemption for criteria, which is about secondary use of identifiable private information. And this brings us right back to those questions that we started with, right? Well, okay, what is identifiable private information on Twitter? Is there such a thing? Or are we going to go with that that perspective that it is de facto public? So I leave us with that to think about these regulatory issues that use use words like when identifiable materials are publicly available. So we're back again, squarely to the public space is Twitter, writ large, a public a public space. And so I want to leave us with those couple of lingering questions about public and private. I'm I think we still have it's it's great. We all this is job security, I think, because we all have so much more to to learn about internet research ethics writ large. And the ethics and the regulations, I don't ever feel that that they're going to be completely in sync. So that puts the onus back on us as researchers operating in an unregulated space to do the right thing. I think it's pretty clear ethics and regulations, you know, aren't always in the same sandbox. So with that, I'm going to turn it over to my colleague, Dr. Marseille. And we're going to talk through some machine learning and some ethics. Okay. So I probably should request control. Yep. And I will just share my screen. Okay. Welcome everyone. My name is Miko Emoże. I work at Poznan University of Technology and the Institute of Computing Science. I work in machine learning. And this is my role, role of me and my colleagues in this project. When asked to present during the seminar, I was thinking about what would be the most because I knew that I would be in the company of ethicist. I would be the only one not understanding what they're talking about. So I've decided to, yeah, to, to at least contribute to the discussion and maybe to demystify machine learning a little bit because people tend to use this phrase machine learning, artificial intelligence, so vaguely, but it's not that complex and it's not that complicated. The engineers can do it. So how hot can it be? So this will be my presentation. I will look at machine learning and what can go wrong when doing research and learning and applying models learned or trained on data mostly harvested from the web, from open repositories, from sources such as Twitter. And we'll, in particular, we'll look at what is, what could go wrong. Now, why do we, why is there so much stress on artificial intelligence or machine learning? Now, if you think about computer science as it is or as it has been for the last 50 years, this is basically it. You take the data, you apply some kind of an algorithm to the data, and you get the results. And whether you are just typing something in your work processor, whether you are computing some equations in your spreadsheet, whether you're browsing in your browser through the internet, this is all the time what you do, right? You have the data either manually created by you or taken from somewhere, you apply an algorithm and an algorithm is just a well-defined finite set of steps that bring you to a desired goal and you obtain the results, right? So the algorithm here had to be written from scratch, had to be programmed. Someone had to write down the code that transforms and processes the data to produce the result. And this is when machine learning comes in because machine learning is indeed a revolution when it comes to ICT. Because in machine learning, this is what we do. We take the data, we show the expected results and the algorithm is the result of what we do. So the method, the steps required to get from data to result is the result of computation itself. And that's why we no longer have to code things manually. We don't have to program them. They program themselves, right? But, and that's why those methods are so data-gritty. The more data you present and the more expected results you present to a machine learning model, the better the model becomes and the more accurate the modeling of the reality or the representation of reality becomes. So this is, I understand that this is quite vague. So let's go into details and let's do some live programming. Consider a very simple example. If you were to teach a child to solve this kind of problem, right? Given three numbers, produce the results. So a child would have to understand the concept of addition and the concept of multiplication and maybe also the concept of which formulas or which operations should be performed first and which should be performed later on. But basically, you would assume some kind of intelligence, some kind of understanding. And this is absolutely not what artificial intelligence does. Remember that we instead of trying to encode the algorithm of solving the problem, we are trying to drive this algorithm by just throwing lots and lots and lots of data onto the machine learning algorithm. And that's exactly how we will solve this problem. We will just create thousands of triplets, just showing three numbers and the result of this computation to a machine. And we will hope that the machine learns how to add and to multiply. So now for coding. This is the only thing that I will do. I will randomly select integers, integer numbers from one to and I will put some caps so not to build two large numbers. And the result will be sum up the first two numbers and multiply by the third number. And this will be the result. So this will be the input and this will be the result. So let's execute this one. And yeah, we are now creating 10,000 examples. Each example consists of 10 numbers. And I have limited the size of or the other. I've put a cap on numbers on integers that we select to be just from zero to 10, just for the sake of simplicity. So let's see how I run it. So this is exactly it, right? This is just the head of this data frame. But as you can see, 9 plus 3 is 12 times 8. It's 96. 8 plus 4 is 12 times 4, 48, and so on, so on. And I have randomly created 10,000 of such examples. And now I will just present those examples. I will keep presenting those examples to a simple neural network. So let's define this simple neural network. And we will run this network for 10 epochs. An epoch is a full scan through the data set. So 10 times the data set will be read by the network. And the network will try to learn how to add and to multiply, mind you, without ever explaining what an addition or multiplication is. So this is our neural network. It consists of three layers. And here is the training. So just give it a second. What you are seeing there is a loss function. The loss function measures the steps of learning. So the smaller the loss, the better the learning. So you can see that we started with some random knowledge or no knowledge at all. And by going through these examples, you see that it goes down, goes down, goes down, goes down. Well, we could continue the learning and probably it would go to much better results. But this should be enough. Let's see the results. This is the test. So this is the set of examples that the network has not seen before. And we will just show those three numbers, ask the network to provide the results, and we'll compare them to expected results. So what we expect to see. So let's do exactly this. And this is our network's response. So sometimes it goes awry. And here is a huge error. We've expected 24, but yeah, the network is totally mistaken. It gives you 21. But other than that, it doesn't look so bad. So after a very, very short time, it learned to add. Maybe I'll repeat this step. The only thing I will change is I will give it a little bit more time to learn. So instead of this, let's say let's give it twice as much time. And now let's generate a new test set and, well, more or less, there is a little bit of error. But still, we have some knowledge. And the subject that interests us the most in the context of this seminar is the bias. So what happens if the training data gets corrupted in some way? And that's exactly what we'll do here. So again, I'll go back to the shorter learning, the number of epochs. But now in each turn, I will change the expected results. This array basically contains the expected values. This is in the training. This is the column that presented the result. So what I will do now is in every 100 elements, we have 10,000 examples. I will modify every 100th example. So I'll just modify 1% of the examples. And instead of having a real number there, I will just say 0. So 1% of data will be slightly corrupted. Instead of containing the true value, it will contain a 0. So you can already see that the training is much worse. The loss function does not go down. And this is, mind you, just a 1% of little error. And let's see our network. Yeah, and the network starts making much larger errors. And especially it starts veering into the negatives, which it shouldn't, because we are adding and multiplying only positive integers. And so it should never produce a negative integer, but it did. What will happen if we do a much more severe modification? So what happens if I input there just 1% of examples? Something very, very large. It's a clear measurement error. So let's see what happens now. And just by looking at the loss function now, you can clearly expect what will happen when I try to apply this model now to the data. So these are the values that we are expecting. So roughly 60, 25, 28, 22, and so on. And here are the responses. The model goes completely nuts. And this is by just modifying a 1% of the data. Probably we could try to do this with one tenth of a percent, and still the network would go crazy for a very, very simple task, right? For just learning how to add and multiply numbers. And this is nothing in compared to the complexity of the task of doing the facial recognition or trying to model the language or trying to model the physicality of the world and so on and so forth. So yeah, there you have it. The bias in machine learning training. Explained as simple as, as simply as possible. So what is this bias? Basically, by the term bias, we mean any type of a systemic distortion of the data. We use the data in machine learning in three different ways. We use it for training the models. We use it for testing the models. So during the model training, we come up with a model. This model can have several hyperparameters, and we're trying to pick the best hyperparameters, the depth of a neural network, the architecture of individual cells in the neural network, the loss function being applied to the neural network. And all these are called hyperparameters. And we can use test sets just to optimize the hyperparameters. And we also need some separate data set for validation. So the data that have never ever been used during the training, and this is the data that we just test the final model on, just to get a glimpse of how this model will work in real life, in the wild, in production, as we say, right? And the bias. So this distortion of the data can be caused by many different sources. Algorithm bias. This is something quite rare, because you'd have to believe that the programmers themselves want to introduce a bias into the code. This is possible, of course. You can think of industrialized Pionage. You can think of just someone being a jerk. Sure, this is not impossible, but given all the pipelines of software production and all the good habits and best practices of software production and code reviews and so on and so forth, this is not very, very likely. Sample bias. Well, this is a very, very significant source of bias. The data may be skewed by the method of capturing. You can rely on historical data. And this historical data, well, it has its own problems due to the fact that they reflect the world as it was 20, 30, 40 years ago. And in many respects, that was not the optimal world that you might want to train your data on. This may be just a stupid way of selecting the data. This may be the survival bias. You only see the things that survived. And these are the only things that you can sample. For instance, if you were to build a model of how well the company will perform in a five years time, and you are taking, for instance, the information from the market from the last 10 years. Most probably you will not see all the hundreds of companies that have failed and cease to exist over those 10 years. You will just see those that survived. So, they will be the best on the market. And you will be training your data on a skewed dataset because it will just show you who has survived and it will not show you the characteristics of all those companies that went down during, for instance, a sudden crisis. This can be as dumb as one of the American cities, which came up with the idea of creating a simple application for an iPhone where it would use the gyroscope in the phone to discover the moment that the phone was moving in the car and it would monitor the trumbles. So, whenever a sudden drop or something would appear, then the application would assume that there is a pothole, right? And that's why the car has suddenly made a movement or something. And it would report the geolocation of the potential pothole to the city services. The problem was that they developed this only for iPhones and there is a clear correlation between your income and, by extension to your ethnicity, to the phone you have. And the city was fixing the potholes, but predominantly in white neighborhoods. That could also be the algorithm bias here. And the measurement bias, some kind of mechanical error, faulty sensor that was, or maybe if the data is being collected by individuals, they bring their own assumptions, their own subjective judgments into the way they record the data, right? So all of that can be the source of bias. Here is a famous example of very, very biased system. It was the model which tries to, and I'm afraid that it is being used. It tries to predict the probability that a person who's seeking for an early release from prison will reoffend. And it's basically, there was a very, very famous study made in 2016, which looked at the, so the problem with this model was that when it was right, it was really right and it was very, very correct. So the precision was high of the model, right? Whenever it made a correct prediction, the prediction was very precise, but when it made an error, a false positive, it made different false positives between different ethnicities, right? So for black defendants, it computed a much, much higher risk of recidivism than actually presented in the real data. And exactly the same thing happened for white defendants who were predicted to pose a lower risk of recidivism than they really did, which came from the records, right? And it was kind of hard to find because the, this bias was present, but only in the part of the model, not in the part of the model, when the model was correct, because then it made exactly the same or exactly precise predictions for white and black defendants, the difference was in other prediction, right? So kind of hard to find and hard to diagnose problem. This is a beautiful example and very relevant to Twitter. In 2018, Microsoft developed an artificial bot, Microsoft Tay, it was called Tay AI and basically they've created a bot, a Twitter bot and they said it will learn from the conversation with real people. So the bot had a language model, it could understand the conversations, it would learn from conversations and they've just given it to the whole Twitter community to talk and have meaningful conversations. And at the very beginning, the very first tweet was see if humans now sleep so many new conversations today. Thank you, so many new beginnings. Now, Microsoft had to pull down the service after 24 hours because this bot has not only become racist, not only misogenic, not only anti-semite, it became an openly Nazi Hitler loving, all due to the fact that people from Reddit and Fortune started doing conversations. And of course, there was an orchestrated effort to swamp the bot with the most offensive and most rude and most terrible conversations one can find in the depth of the internet. But yeah, these are the tweets, tweets generated by the bot after just 24 hours of having conversations with humans. That speaks more to the nature and the state of humankind than to the progress of Microsoft's engineers. But anyway, it cannot really depend on the user-generated contents, especially when the users have an agent with respect to your AI. But it doesn't have to be so malicious. Here you have the location of Google's office in Berlin. And as you can see here is a terrible traffic jam. This is the street during the traffic jam. This gentleman here walking, he created the traffic jam. What he did, you see this small little trolley. This trolley was loaded with 100 active mobile phones. He was just walking the streets with those phones. And Google Maps was recording the location of all those phones and seeing that those phones are really slowly moving. So assuming that this is a terrible, terrible congestion on the streets, it probably suggested everyone else to just go somewhere else and to direct the movement to nearby streets, as you can see. So the guy had the street for himself. This is an example of, of course, this is not malicious. This is benagaland. But still, you can see how a service which is very sophisticated, very complex, very large involving hundreds and hundreds of very skilled engineers can be fooled by a guy with a small trolley and a couple of bucks to spend on, on phones, or just asking his friends to borrow phones for, for 15 minutes, right? So yeah, this can happen as well. The bias, the bias can be algorithmic and can, can be created by humans. This is another infamous example of the Project Greenlight in Detroit. So here you see the locations of CCTV cameras across the city. And here is the distribution of ethnicities in the city of, of Detroit. And it's really hard not to see a very, very certain pattern of placing those cameras. And of course, the placement of cameras. So the placement of sensors directly influences the selection of data, right? Because you will get the data that you get. In other words, if you see, for instance, if you try to use those cameras, for instance, for measuring or say those cameras can measure the speed of the car, they will learn that only people of specific ethnicity break the speed limits in the city, not because the only reason will be that this, the model trained on this data will not see other faces, right? So, and you can imagine that the data collected, the moment of data collection, or even worse, the model of, the moment of model creation is postponed by several minutes, several years or months from the date of the selection of places where the cameras are. And then people take the location of cameras for granted and they don't question that. They just say, okay, we have the feed from the cameras. Great. So let's pull the pictures and let's train our artificial intelligence models to do this and that and that, right? But you have to go down many, many years to see where the cameras were located, why they were located, where they were and think about what might cause, what damage might be caused by such selection of places. Similarity bias, again, something that is present, very present in contemporary machine learning models. This is something that leads to information bubbles, right? If you search Google News for an article and you give it some keywords, it will find articles and it will find other articles with similar headlines. And the headlines, given a very specific selection of keywords, they will mostly corroborate a given point of view because the same facts can be reported totally differently and with different keywords, different speaking points, depending on the political affiliation of news source. So if you're just using the recommender engine and you search by similarity saying, yeah, the person wants to read this, so let's recommend more similar news, but similar in what sense? The similar in terms of the subject or the similar in terms of the form? If the latter, then probably it just enforces the information bubble because it will show people exactly the same point of view. YouTube has chosen a terrible objective function to optimize for the total length a person spends in the service, not on the number of videos being shown, not on the number of ads being shown, not at the quality of videos being shown, not even the similarity of videos descriptions, right? They were just looking, they were optimizing the recommender system just to keep you as long as possible in the feed and as a side result and nobody programmed that. As a side result, this promoted a huge amount of extreme content and all kinds of conspiracy theories being displayed in those videos. Or some unforeseen consequences of aligning with stereotypes if job adverts are presented to people and they present, for instance, say medical technician versus a nurse and someone, a woman would or could, for instance, select the nurse just by self aligning with a stereotype of a woman doing the job of medical technician is a nurse, right? So that is quite a problem. Under representation, these are two extreme examples, but I just wanted to share, I was thinking whether I should show you those pictures or not because they are very offensive, but yeah, this happens in products produced by world leading manufacturers. This is a Nikon camera, right? And yeah, it is supposed to be a helpful tool to suggest you to repeat the photo if someone has blinked, right? And they haven't trained the algorithm for discovering people blinking on the Asian people, right? And they are assuming that this is this person blinking. Come on, it's really hard to be more offensive than that or not to mention this one. This was so famous that this is borderline criminal. Care to guess what it is? These are the results of a Bink image search by, if you just type in CEO into the Bink image search, this is what you will see. How many women do you see there? One. What percentage of American companies are being led by women? 28 percent. And 28 percent of CEO positions in large companies are being currently held by women. So this result from Bink images is really something to behold. But not only images, just let's play a little bit with text. I wanted to write something and yeah, we'll go from Polish to English to French to Turkish and back to Polish. So I will write now in Polish the phrase, she is a famous actress. And in French, because French also has a grammar gender, it is une actrice célèbre. So this is feminine, right? Okay. So now we'll go to a language that doesn't have a grammatical gender. Turkish language is an example of such language. So we go into Turkish. So now I will translate, I will switch the translation from Turkish to Polish. And now it says is a famous actor, but actor has already a masculine ending. So it took a masculine grammatical gender. So let's go back to Polish to French. And at un acteur célèbre. Before we had an une actrice célèbre. And now we have an un acteur célèbre, right? So woman gets lost on the way because a famous actor must be a man. As a matter of fact, the problem is with the translation to Turkish where you drop the gender, you drop the gender, grammatical gender. But then if you go back, you have to either reconstruct it or you should produce at least two different versions, right? And not just one. Yeah. And to close my presentation, after all this, let's play a little game. Are you the source of bias? Just look at those images and imagine that you are a human annotator who is responsible for providing labels for an machine learning machine, a machine learning algorithm, machine learning task to teach a machine to automatically label images. One of those three descriptions is seriously wrong. Can you spot which one? I'll give you just a second to think about it. Or maybe someone wants to propose, wants to propose the label, which is clearly wrong. I don't, I don't see anyone. So I will tell you which one. And of course, it is a black woman place with her daughter. And the problem is that it is not a black woman. It is a woman. The adjective black has nothing to do with this image. The only reason why you would like to inform a machine learning model that she is black was to contrast her with someone else on this photo who would be, for instance, white. So a black woman talks to her white colleague. That would make sense because then the adjective would help the model to recognize between her given her skin complexion and the colleague who would have a lighter skin complexion. You don't see here a white man placed with a dog, right? Because you're assuming that he's a man. So he is a man and her color, the color of her skin in this labeling of this image of this action, what she does with the kid has absolutely nothing to do with the action, right? And this is very, very hard to spot, especially if you're not a person of color, to spot that this adjective is not only superfluous. It is wrong because it teaches something, the model that it should not teach, like this adjective serves some purpose, and it serves no purpose in this label. So it's really much, much harder than one would think. Okay. Thank you very much. Okay. Thank you, Mikolaja. That was fascinating. We are going to turn it over now to Jan, and Jan's talk is entitled, An Ethical Framework for Web Immunization Score on Twitter. Yes, I'm just trying how it goes. Please give me a second. That it works right now, that you can see my presentation, and hear my voice when I'm speaking. So, yes, I'm going to talk about our research project, and yes, excuse me, because I become a little bit distracted when I try to start the presentation. So I'm going to talk about ethical framework for a web immunization score on Twitter. And we defined web immunization as individual or group susceptibility to misinformation on social media. And this machine learning element of our project is only part of this project, and I will try to focus only on this part and try to analyze ethical elements of this project. But at the beginning, of course, I have to make some disclaimer and disclosure. I have a conflict of interest. However, this is not a financial conflict of interest. This conflict is due to my double role in this project. So on the one hand, I'm a project leader. So of course, my goal is to navigate our research project to the fruitful end. And I would like to omit and avoid all possible problems. But on the other hand, I'm a bioethicist interested in research ethics. We are background in philosophy. And I would like to analyze all important ethical problems posed by our research project. And we invite you to take part in the seminar. We also invited some external experts also to avoid our blind spots to really take advantage of some outside perspective. I'm going to analyze our project, and I will be using ethical framework, which was elaborated by Norwegian National Ethics Committee. It is present in the guide to internet research ethics. And this framework consists of four different dimensions that should be covered by every ethical analysis. So the first dimension is accessibility of public sphere. The second is interaction with participants. The third is sensitivity of information that is collected during the research project. And finally, the fourth is about vulnerability of participants. So let's get started with the first dimension accessibility of public sphere. So it was already mentioned by Elizabeth. Twitter is usually considered to be a public sphere. And the goal of our project is to build machine learning models that will predict individual and group web immunization scores. So individual and group susceptibility to misinformation. And it will be based on their activity on social media. But in our project, we select actually to collect data from one social medium from Twitter. So we will use Twitter API services to collect massive amount of data and be able to estimate individual and group susceptibility to misinformation. And of course, the first question that can be asked is, do us researchers, do we have a right to collect identifiable data? And how our right can be related to users' expectations? And the data on Twitter is very difficult to be de-identified or anonymized. Because every single tweet is connected to the whole conversation. And in a tweet, not only the content of the tweet is important, but all metadata that is associated and linked to that tweet. So in order to really understand the tweet, we have to connect it to other tweets to the whole conversation. Is it just a standing-alone tweet response to someone? Red tweet with quote, who send this tweet and mapping and trying to put a single tweet into this whole context make it almost impossible to de-identify. Because if we identify a tweet, it loses its research or data potential. It becomes useless from researchers' perspective. And generally, as it was mentioned by Elizabeth from the regulatory point of view, of course, we are allowed to do both terms of service and development agreement of Twitter allows us to collect the data and also terms of services on Twitter. They are very explicitly say that Twitter is disseminating the content that users are sending on Twitter. However, the developer agreement puts some restrictions on how this data could be used. I will discuss it a little bit later. From the federal regulatory point of view, so both from European perspective and GDPR and from US perspective, from the common rule, this data is considered to be more or less public and available for research. And for instance, Article 9 of GDPR says that it is not prohibited to process data which are manifestly made public. And one can reasonably argue that the data on Twitter are manifestly made public. However, as Nicholas Gold realizes, Twitter imposes certain restrictions on researchers and those who want to use the data from Twitter. And for instance, Twitter forbids to reuse deleted content. And that's why Nicholas Gold says that we should not consider data on Twitter to be public data, but rather private data on public display. And we can also see that Twitter users are even not aware of the fact that the majority of Twitter users is not aware that researchers use Twitter for research sake. And however, quite substantial percentage of Twitter users would not opt out from research if it is possible. So they would still participate in research. But one third of research users, if it were possible, would opt out from research and would not like to provide their data to researchers. So we are in the situation when we have to somehow balance the discourse of data ownership, which is mostly present in the US context. And in the European context, we rather think about data as an inalienable individual possession, which can be controlled by individuals and political community. But we have to also balance these two aspects with public benefit. So on the one hand, this is private data on public display. So we should also respect its private or this element of control of individual control, but also weighted against possible public benefit. And when we talk about another aspect of research, interaction with participants, at this stage of our project, we are not going to interact with participants. But sensitivity of information, however, is very important in this context. And especially important is the concept of group privacy. So usually we think about privacy in terms of individual or group rights. And we even think that well-defined and self-proclaimed groups, such as families, ethnic minorities, or even group of patients who are diagnosed with specific conditions, that they have a certain rights and they can claim these rights and seek justice in front of a court. And to give an example of such group claim, we can, for instance, think about Havasupai tribe. That is, in bioethical context, this research project was quite popular to discuss. So researchers violated, let's say, privacy of this ethnic minority of this ethnic group because they used their blood samples without community consent. And they assessed risk of mental disorders, such as schizophrenia and alcoholism. And they also used their genetic material to study their history and the genetic origin and doing so, they undermined their self-identity beliefs. And the tribe recognized it as a violation of their group privacy and they sued the university. But when we think about group privacy in the context of machine learning, we cannot use this concept of well-defined group, which can have certain legal representation. Because these groups, which are formed in the process of machine learning, when we discover certain characteristics and we can say that one individual belongs to that group, one individual, by the way, can belong to many different groups, these individuals are not even aware of this fact. They don't know that they belong to one or many different groups. And they don't have any kind of representation, but still they can be a subject of certain algorithmic intervention. And because they don't have knowledge about this intervention, they cannot seek redress before the courts and they are not recognized by the legal system. And one thing that has to be mentioned also that the concept of group privacy is very closely related to profiling. And Twitter developer agreement explicitly prohibits profiling. So the Twitter developer agreement says that targeting, segmenting, or profiling individuals based on sensitive personal information, like health, negative financial situation, and so on, cannot be used by those who use Twitter API. So right now we have to ask, we are facing three, let's say, ethical questions. So the first question is how we should treat Twitter development agreement. Is this agreement legally binding or ethically binding for us? Do we really exhaust and meet the definition of profiling in our research project? And how we are going to protect group privacy? How we are thinking about protection of group privacy? So the first question, how important is from ethical and from legal perspective, Twitter agreement with a developer? So from the legal perspective, that there is already some case law, which indicates that this kind of agreements are recognized, at least by the American justice system. However, I think that there are quite strong ethical arguments which can say that in certain circumstances, this agreement can be, say, violated or overridden. So companies such Twitter, Google, Facebook have a very strong influence on our politics, not only on elections, but also on discourse and political discussions. And I think, in my opinion, a democratic society has to have some instruments, have to oversee their actions. And researchers are probably best situated to really put a check on these companies and examine their activities. However, of course, this kind of research projects should be carefully overseen and reviewed by external ethics committees and also should have legal support from research institutions. And of course, right now, I want to stress and emphasize that we are not going to violate Twitter agreement. And this is not what we are thinking we are doing in our research project. So what is the definition of profiling? So profiling is a technique to automatically process personal and non-personal data aim at developing predictive knowledge. And that knowledge should subsequently be applied as a basis for decision making. So I think that we have to draw attention to these two elements. So on the one hand, predictive knowledge, on the other hand, decision making. And of course, our project aim is to create predictive knowledge. But we are not going to make any intervention about these individuals, which, let's say, provides us with the data. However, we are very aware of the fact that this data could be used in that way. However, we won't perform any kind of intervention at this stage of the project and every future intervention would be coupled with informed consent process. What about vulnerability of participants? So the last dimension of our analysis. And vulnerability, usually in the context of biomedical research, refers to the moment when we involve participants into a research project. And those who have some kind of cognitive and those people who do not have sufficient cognitive capability or who don't have legal capacity are usually unable to make informed decisions about themselves. And they are recognized to be vulnerable in the context of biomedical research. But I think that this concept of vulnerability really doesn't apply to our research project, also because we are not going to obtain informed consent from our participants, from the Twitter users. However, I think that some of our participants are vulnerable in that situation because I think that susceptibility to misinformation can be understood in terms of vulnerability. So people who have diminished ability to make autonomous decisions in the information environment of social media are vulnerable in these circumstances. And those who cannot recognize misinformation and who spread this misinformation are vulnerable in information environment. So what about protection? I said that informed consent in our project would be very impractical and we are also from the regulatory perspective, it's not required. However, Neil Dickard and his colleagues in a very interesting article published in American Journal of Bioethics recognize that informed consent is a procedure that has a lot of different functions. Actually, they distinguish between seven different functions of informed consent. And this function can be also realized by other procedures or other actions. So for instance, one of the functions of informed consent is to make the process of research transparent. And by being present on social media and also disseminating information on our research project, we are going to, let's say, to try to meet this element, this function of informed consent. So we would like to inform, let's say, Twitter's sphere that we are conducting this kind of research and what it means for Twitter users. Another form of protection that we are thinking about is limited data sharing, especially when we talk about this data set that contains data from Twitter and the model that will allow us to predict the web immunization score of individual and groups, because we don't want we don't want this model to be used by any vet actors to, for instance, target susceptible individuals and groups with this information. So we were thinking about some kind of data access committee that will limit and vet the data request. And also we will not share the very model of data, but only a surrogate model, which allows to validate our research, but which doesn't allow to, for instance, to replicate and to target vulnerable individuals. And generally, I think that our research project, of course, there is a lot of ethically, let's say, sensitive issues. I think that we rather that one of the ethical loss of information ethics, which was formulated by Florida, which says that entropy at not be caused in the infosphere is justification for our project. We want to limit misinformation and this is also the how we serve, let's say, public interest doing this research. So we are not doing this for just for fun or just out of pure curiosity. Okay, thank you very much. Okay, thank you, Jan. We're going to turn it over to Nick Preferis. One second. All right. Can you see my screen? Yes, we're good. Okay, excellent. So thank you all very much for having me here today. It's a real pleasure to be here to talk with this group. And thank you very much for the opportunity to also reconnect with Dr. Picanon. Dr. Picanon, of course, was one of my very early mentors in my PhD. So it's a real pleasure to be here. And hopefully I do her proud. So I'm going to talk today about Twitter research ethics and thinking about ethics sort of beyond the review ward. And my talk is broken down into three parts. So I'm going to move essentially from quite broad into sort of a specific example. So I'm going to start today by talking about ethics very generally what we mean by ethics and the use of publicly available data in relationship to these different meanings. Next, I'm going to talk about some research that I've been involved in in terms of understanding how researchers themselves are going about their ethical practices when using data from Twitter and talking about a little bit about the gaps between what researchers do and what users actually think happens to their data and some of the problems they're in. And then third, I'm going to talk about a very specific example of values tensions that exist in relationship to data sharing around Twitter and some of the tensions between what we're sort of required to do by rules and regulations and what our ethics may demand and how we might be able to go about practically solving some of these issues. So I want to start today by thinking really, really big about what is ethics, why do we care about the stuff? And, you know, when I teach my undergraduate students in my information technology class about ethics, you know, I go back to this idea, this very old idea that's been with us very, very long time, that ethics are systems of principles that we use to guide us in making moral evaluations and we can rely on things like utilitarianism to help guide us in terms of making decisions and evaluating decisions based on what's going to provide the maximum good. We can look at something like content and deontology, which is going to suggest that we should evaluate a given action based on what our duties are in that particular circumstance. And we actually might ask not just about act evaluation, but also about character ethics, virtue ethics, for example, is going to ask us and help us reflect on the kinds of people that we want to be. Are we instilling the kinds of values in ourselves and in our actions that are virtuous? And at their core, these ethical principles are really in systems are really about using our capacities for reason, judgment, and thought to critically examine our own actions and our own character. And so I really want us to keep this idea in mind that ethics is about in part using our capacities for reason and judgment to think through our actions. Now as researchers, we often talk maybe colloquially about ethics as sort of that regulatory piece, that compliance side of things. Maybe you're having a hallway conversation with a colleague and you say, oh yeah, I have to do the the ethics piece of my project now, right, going through the IRB, getting the paperwork for informed consent approved. Certainly I go through this, I'm sure all of you go through this with your work as well. This is about a process of ensuring conformity with relevant laws, policies, and guidelines. And of course, these laws, guidelines, and policies are developed in relationship to particular ethical principles that, for example, Dr. Buchanan mentioned at the very beginning, respecting human dignity and autonomy, maximizing benefit, minimizing harm, so you can see the trace there of utilitarianism thought in that particular value, justice and beneficence. Now ethics as compliance is really about ensuring that researchers don't violate certain sort of baseline conditions for the treatment of others essentially, right. These were set up to make sure that researchers do not violate what I would call flooring level needs to make sure that people don't violate the basic conditions for how we should behave towards one another. That doesn't necessarily mean that that is the end of our ethics though. And in fact, compliance and contemplation may not necessarily be playing in the same sandbox as Dr. Buchanan reminded us. Sometimes, compliance doesn't cover ethical situations that we may encounter as part of our scientific research practices. And very often, when using public data from social media sites, we're getting more and more, you know, we're getting better policies today than we had say a decade ago, but there's still incredible gaps where the rules and policies that were developed for the biomedical setting may not apply cleanly to the context of the social media using public data. And part of this is because at least in the U.S., and I'm not an expert on EU policy, so I have to apologize there, but in the U.S. at least, public data is often not considered to be, the use of public data is often not considered to be research involving human subjects and therefore not subject to the kinds of oversight of institutional review boards. But again, this is not to say that these research projects lack ethical components. It's simply to say that they fall out of the purview sometimes of the compliance side of ethics. And I want to make the argument here that ethics and our ethical reflections and our actions and evaluation of our actions needs to be considered across the entire process of doing research. And certainly, there are, in fact, a litany of ethical quack miners that we may encounter. These can happen during the data collection process, the data use process, and the data sharing process, so the entire gamut of the research process essentially. So certainly in terms of data collection, we might think about, for example, data that was perhaps public once, but it's now been deleted. There may be ethical questions about how we should treat that data. We might be using the data from marginalized groups. It may be public data, but we know that there has been overburdening of particular populations. And so even when we're using the public data, we still may have obligations because of the historical injustices on those groups. Aggregation of data points to create a very detailed picture of someone's life can be a threat to someone's privacy and may have ethical dimensions. And there's questions that we should ask about, even when it is public data, if someone contacts you and says, I would like my data removed from your data set, whether or not we should honor that request. As part of data use, we might encounter questions about the actual ends of our project. So a really interesting example actually in just the past couple of weeks, researchers at the University of Minnesota had a project that fell out of the compliance side of oversight, where it was not considered to be research involving human subjects, but where they were purposely introducing errors into the Linux kernel in order to study whether or not the development community would actually find those errors. So we can recognize that there's some serious ethical issues about, for example, purposely introducing errors into a system that thousands upon thousands of people rely on, even if the IRB says, oh, it's not research involving human subjects review. And sometimes we also see, you know, I also like to use the example sometimes of projects that scoop up lots of public image set data to do things like use images, pictures of people's faces to try to predict the political leanings or try to predict sexuality. And some people have labeled this as sort of phrenology to point out, right, obviously has severe ethical implications, even though it's public data. An important thing that's actually come up really recently that I do want to particularly encourage this group to think about as they're studying misinformation and disinformation that's been actually that's come up in the context of research around Gamergate is not just thinking about your obligations towards your research subjects, but also thinking about your obligations to your fellow researchers and to, for example, your students. If you are studying by Trollic content, content that could be psychologically harmful, you know, that maybe it's violent information, maybe it is disparaging information, finding ways to make sure that students or your fellow researchers have the kinds of support mechanisms they need to be able to engage with this work and have, you know, essentially support if they're feeling harm themselves by the kinds of content that you're analyzing. And then finally, in terms of data sharing, there's some really serious ethical questions around how we represent our data subjects. So for example, the power of labeling. So an example might be something like if we create a data set and we label this data set tweets from people that we think have depression and then make this publicly available, right, the way that we're representing that data subject, those individuals can have implications for them. And so, you know, we have to be reflective of the implications of how we represent our subjects. Certainly, in terms of data sharing, we also want to think about things like other ways of sharing our research outputs with the communities that we're actually studying. This can, you know, obviously have a lot of benefits for the researchers in terms of developing connections, but can also be a question around justice. And then finally, you know, certainly when we're thinking about upholding values for science, replicability is a really big ethical issue. We want to make sure that the work we're doing is valid, that it's reproducing, and that we can benefit science and human knowledge going forward. So the thing I really want to emphasize in this first part is that there's no singular ethics portion of the research project, right? It's not just something that you do upfront. It's a process of continuous evaluation of our actions, of our character, where we weigh our values and duties. And I'll point very briefly to a framework that I think is very useful in helping us go through that sort of continuous process of evaluation about maybe conflicting values or duties or values and duties that are in tension or tensions between what we're asked from a compliance side and what we may feel from a, you know, contemplative side. And that is the process of using disclosive ethics. This is a framework put forward by Phillip Bray that asks us to sort of continually engage in a descriptive process of what the values tensions are or where there's gaps that we're noticing between different values and a normative component, how we then go about actually addressing those gaps, one which we have described them in full. So, I want to pivot now to getting a little bit more specific. So, that was sort of my broad introduction to thinking about ethics and questions in relationship to the use of public data. I want to talk now about Twitter specifically. So, obviously, Twitter has become a major source for academics. There's been over 2,000 research papers published in the past three years using Twitter data. Projects, obviously, include sometimes billions of tweets now. And we certainly have a reliance on Twitter, not just because it's become a very dominant space, you know, for political discourse, for responding to events, but because also it is easy to get the data there, because relatively speaking. And because the data that we can actually get is a kind of data that we can digest very easily. It's kind of data that is very easily parsable for machine learning applications. Textual data is obviously much easier to process than image-based data, other kinds of video content, and so into make inferences from it. And because comparatively, it is relatively public. Now, I have public there and scare quotes, and I'll unpack that a little bit in just a minute. And I want to make note that as part of a research project I did with Michael Zimmer about five, six years ago, we looked at research that had published using data from Twitter and found very few of these projects actually report about going through ethics review. Now, that isn't to say that they didn't, but only about four percent of the published research that we were able to find actually talked about going through IRB or talked about their ethics processes. And I actually do want to emphasize that I do think it's quite important that we talk about our ethics, that we include it as part of publications and include that sort of descriptive component and our evaluation of it. So I want to talk for just a second about a research project that I was involved with, with Dr. Casey Feisler at UC Boulder, where we sort of recognize that Twitter data is becoming more and more prominent in terms of its use in the academic setting. And we kind of had the question, well, what's the other side think, right? Do users actually know their content is being used for academic study and how do they feel about it? So we surveyed Twitter users asking users whether or not they think that researchers are allowed to use their content without having to get reconcent. And then we started asking them questions about their level of comfort with the idea of their tweets being used for research. And in particular, we were trying to understand sort of the contextual factors that might drive users level of comfort with the idea of their content being used or used in different ways, such as whether or not they were asked for permission by the researcher, the kind of study it was, the kinds of content that were being analyzed, and whether or not they were, for example, quoted directly in the study or indirectly. So the top line statistics that I thought were really striking were that we found over 60 percent of our respondents actually thought researchers were forbidden by Twitter's terms of service from using public tweets without having to ask the users for their permission. That is not the case. This is an incorrect understanding. Now we asked, we told them after this question that actually they are allowed to do it, but would you like them to get your permission? And 65 roughly percent of our respondents thought that researchers shouldn't be allowed to use tweets without permission without having to go back and ask the user. However, and this is a thing I do want to emphasize, that many users are actually somewhat comfortable with the idea of their content being used if asked, if they can see the research outputs, but this is extremely contextually driven. So I'm not going to sit here and read this entire chart for you, but I want to emphasize and just try to give you a quick read of what's going on here. So these are the different contextual factors that we asked about on the left hand side at the top. This is a essentially a Likert scale between very uncomfortable on the left hand side here and on the right hand side, very comfortable and scores in darker blue are higher prominence. So for example, users are very uncomfortable with the idea of their their content being used and never being told about it. Researchers, excuse me, users are very uncomfortable. Over 50 percent indicated they would be very uncomfortable with someone using their tweets from a protected account, which is a kind of privacy that you can invoke on Twitter that controls the distribution of your tweet. People were very uncomfortable with the idea of someone, a researcher using a tweet that you'd created, even if it was public, but that you had later deleted. Now, where would you see increasing levels of comfort? Actually, and I think this is kind of interesting, is the idea that if my tweets are being used in a big data set, maybe I'm a little bit more okay with that. So for example, people were much more comfortable with the idea of if I was just one person in the scope of a research study that's studying a billion tweets, I'm a little bit more comfortable with that than a situation in which a researcher is only studying a few dozen tweets or a few dozen people. Interestingly, users are much more comfortable with the idea of an algorithm analyzing their tweets rather than a human, which I think is kind of interesting. Folks are very uncomfortable with the idea of researchers using not just your tweets, but also other information in tandem with tweets. So for example, public profile information such as location username. And folks are pretty uncomfortable with the idea of being quoted in a research study with their Twitter handle attributed to that quote. They're a little bit more comfortable with the idea of their tweets being used in studies if they're attributed anonymously. So there's some interesting contextual findings here, but the big sort of things I want to draw our attention to are what I think are potentially tensions in some of these results. So users' understandings of how researchers are actually using their data are pretty limited in my view based on this data. But they do have contextually driven levels of comfort, which I think are important to acknowledge and to understand their perspective. At the same time that we recognize from this data that there's some serious gaps in users' understanding, it's also important for us as researchers to promote certainly the progress of science and values that sort of again serve that wider public of addressing misinformation and abuse that are happening at a much more macro level. And one of the things that people have talked about, for example, is, well, is there a way that we could perhaps think about notifying users when we're using their data for research? But in some situations, actually, that notification process itself could cause users more anxiety that they're being watched or being studied than actually no notification. And so I just want to point out here that there's different tensions at play and it's a really important to describe this space in order to understand these tensions. So again, so I'm moving now from really broad talking about ethics and public data to talking about Twitter to talking now about a very specific context, which is data sharing and exploring some of the tensions in relationship to data sharing on Twitter. So obviously in terms of the progress of science, a very important part of upholding scientific truth is ensuring the validity of our work that's part of and also ensuring public trust. A big way that we do that is by providing our data and providing our methods in order to try to create opportunities for replicability studies. And this is really important. Obviously there's been a lot of concern about a crisis of replicability in scientific research, particularly in the social sciences in the past decade or two. Now, in this particular situation, thinking about the context of Twitter data, one way that we can enhance the sort of benefit to the broader community scientific community is by making our data sets available for use by others. But there are some tensions here, obviously on the compliance side of thing. Twitter's terms of service forbids us from sharing the full JSON data. So this is the full data of a tweet. So we would actually be essentially told that we cannot actually provide the content of a tweet in terms of resharing certain metadata fields we're not allowed to share. Twitter's terms of service actually only allows for researchers to share with third parties of the tweet IDs themselves. Further, our local laws may or may not allow resharing of data. And at the same time, we also might be asked by our funders. So for example, National Science Foundation might ask me to share my data as part of trying to ensure open science practices. So we already have tensions just in the compliance side of things. From the sort of ethical contemplation side of things, we realize that users themselves may not want to be in a labeled data set. Again, think about a labeled data set of tweets that we think are from people who, based on the sentiment of their tweets, we think may be prone to depression. Labeling someone in this way could have implications for their privacy if I can look up a tweet ID and then be able to see the user that it came from. We also might ask about whether or not there's an even distribution of risk that's being shared among the population in relationship to the population that we're studying. Again, thinking back to situations in which there's marginalized groups that have historically already been overburdened or have had injustices committed against them in relationship to scientific research. And then we also, from the contemplative side of things, also have to think about, well, we still want to uphold open science, the validity of our work. What happens if data starts disappearing from this data set? A researcher at the University of Maryland at Summers did a project not too long ago where he looked to see how much content from a data set had been deleted after a year, after an event, and found that somewhere between 10 and 15% of his data set was gone, essentially, where users had simply just deleted that content after the event occurred. And that really threatens our ability to recreate an event, to understand a particular social phenomenon. So we have a lot of different tensions here. It's really important for us to describe these tensions and think about ways that we might go about solving these tensions. So I want to draw your attention to one project, the Documenting the Now project, which has developed a set of code that they call the hydrator that lets researchers easily rehydrate, grabbing all of that extra JSON data from Twitter's APIs, simply based on a list of tweet IDs. So this is a piece of code that helps sort of address the fact that Twitter's Terms of Service requires us to do one thing, but that we may feel that is in tension with questions around difficulty of actually going about doing this process. Certainly, we want to think about, in terms of data sharing, finding ways to anonymize user data when possible, particularly if it's potentially sensitive content or if we're doing a kind of sensitive labeling of a data set. We might think about ways of addressing these tensions by doing things like making our data sets available only by request or setting up specific kinds of resharing agreements. So there is a series of data sets called the e-risk data sets, which is a set of data sets used for developing better machine learning models for doing things like predicting self-harm, predicting depression. And these, this is a set of tweets from individuals that have been, have a baseline where they have indicated that they have had a positive, for example, diagnosis of a particular mental state or have had a positive case of self-harm. And essentially that's the ground truth that they're operating against. And so this data set, you can get access to it, but you can only get access to it by request. And there's a specific user agreement that the researchers have set up with terms to limit what researchers can actually then go and do with this data set. And the data set itself has been scrubbed already of as many identifiers as they can possibly, they think realistically scrub from it. But the terms of service that the researchers have set up for reuse also includes terms like making sure that future uses can't try to positively identify you're forbidden from trying to re-identify the individuals from whom this content originated, you're forbidden from re-sharing this. So thinking about ways of setting up your own terms for data sharing is one way of trying to address the tensions. And then finally, the thing I want to come back to is document your decision-making, discuss your decision-making. One of the big difficulties of coming up with ethical standards in this space is the fact that very often we don't report our ethical thinking and our ethical decision-making as part of our publication practices. Getting to a point where we talk about our norms and normalize them is critically important for this space. So again, our ethical evaluation needs to be a continuous process through the entire research process from A to Z. And we need to document that and discuss it. I will leave with this note that certainly the Association of Internet Researchers has been thinking about this particular ethical questions involving internet research. They've used public data for decades and they've published two ethical decision-making guides that are available online. The thing I really like about these guides is it's not a compliant side model that if you just do these things that you will be okay. It really is a series of prompts that are set up to help us think through the particular challenges that we might have as part of doing our practices. So it's really there to help us think out potential situations. And so with that, I'd like to say thank you. That's great. Thank you, Nick. Thank you, Mikalaj. Thank you, Jan. We do have time for some questions. Let me just check the chat real quick. I see, hi, Natasha. I see your question. I wanted to ask about the expectations of internet users. Expectations of internet users, the public visibility of online research projects and disconnect between users' expectations and the reality of online data collection and use. Do the project team see it as a part of their ethical responsibility to try and educate internet users about what is happening to their data? And how might this be achieved when consent is not being sought without alienating users, as Nick just said? So Jan, do you want to start with that and then we'll move across? Yes. I'm thinking about this question. I was a little bit, again, distracted with my video, so I hope that you can hear me. Yes, I was thinking, of course, about internet users' expectations. I was also thinking about their expectations from a little bit of a different point of view, because for a while, I was very much focused on ethical aspects of using electronic health records. And there is already a lot of research, especially focus groups with patients who are asked about their attitudes to using their electronic health records in research. And usually people at the beginning of the conversation, they're not feeling very comfortable with research use of their health records. But when they are informed about possible health benefits and about all privacy protections, then there became more and more willingness to share the electronic health data, especially if they have a certain condition. So I was always very curious about, let's say, if we are asking a certain questions, just, and we are using certain instruments to measure users' attitudes, sometimes we will not, let's say, help users to really evaluate their answers. So of course, I take this into consideration, and it's a real concern that people are really reluctant about sharing their data, and they treat research with a huge suspicion. And what we, of course, what we want to do, we want to be as much as possible, as transparent as it is possible. And also, this kind of seminar is our effort to make our research project visible. And we will be, of course, informing and explaining. And here I see that there is also a part of question about consent that, let's say, it's impossible because we have to contact hundreds of thousands of users, and that will be impossible to physically process. We don't have time and resources to really provide this information. And mostly when people are asked to participate in this kind of research, they can even consider, for instance, asking them as a form of intrusion. Then they don't have any obligation to respond to our invitation, to even read the informed consent. So from the organizational point of view, it is impossible to really conduct research with machine learning and asking for informed consent. Mikko, maybe you can also add something about this. Oh, not really. I think this covers basically your task. I really, first of all, I don't think we are touching upon such a private and directly vulnerable feature of people. Of course, their vulnerability to fake news is a vulnerability, but it's not a direct vulnerability that can be very easily used by an adversary. Besides, we already had discussions about how to present the results of the project. For instance, we both with Jan agreed that probably publishing all the models that we will be developing is not a good idea, that maybe we will just publish a surrogate models, so models of models instead of detailed ordinary models that would allow you to score a given individual against the vulnerability. So yes, this is an ongoing and fascinating, and for me especially a fascinating discussion, because this is the first time in my life that I'm having those discussions. It's not something that in the technical and ICT communities is at least not to a recent, not until recently, this conversation has not been going on very much. And now it is, we're catching up with the rest of the civilized worlds regarding the ethics, so hopefully we will catch you guys. Yeah, just to respond back to Natasha's question too, in the US, in the regulatory framework, we actually have a word that we use that getting consent from that many individuals would be impracticable, and so it sets it outside of the parameters of a typical informed consent process. Okay, there's another question. I can read it during the pandemic, many researchers performed ad hoc analyses of Twitter without any ethical consideration who just entered this area of research recently. Could you talk something about this phenomena to prepare a study protocol? You need a lot of time, so especially in the beginning of the pandemic, some investigations could violate some standards. Does anybody have any specific examples of that happening? I don't know any specific example that I can discuss, but let's say usually when we talk about normal biomedical research during this kind of emergency situations, usually the review protocol is performed in a very, let's say, expedient way. So it is shortened and it is usually not a full review. Sometimes even this kind of protocols are reviewed in advance, so they just wait for the pandemic to be launched and reviewed. And I have to admit, and this is also why we are organizing this seminar, that online research on Twitter is also very new for me, so I'm also learning what are the specific ethical standards for Twitter research. Because as it was mentioned before, from the regulatory point of view, researchers do not violate any specific regulations. And even from the perspective of bio-researchers, all these restrictions that could be imposed by additional reviews or consent forms or something like that could even seem to be a little bit excessive. Because right now there is a discussion within, let's say, bioethical community, how to facilitate ethical reviews, how to loosen ethical restriction and allow researchers for self-regulation. And as Elizabeth mentioned, Common Rule was quite recently upgraded and revised. And the expectations from the biomedical researchers were that even that the self-regulating aspect will be taken into consideration even to the greater extent that it was. Because right now, in bioethics and in biomedical research ethics, we discuss rather, and this is a phrase taken directly from the article written by Tom Beechamp that over-protection, over-regulation leads, especially in this kind of studies of emergency pandemic studies, to under-protection. So over-regulation and over-protection leads actually to under-protection. But I'm not sure if we can use exactly the same logic to internet research. And I think that generally internet researchers are in a very, right now, in a very nice position because I don't think that there is any pressure from the regulatory point of view to really tighten the regulation. But they have a, I think that they have an opportunity to self-regulate themselves and to set this ethical standard by themselves, also in order to build trust with participants, with users, with those who produce data. I don't know what is your opinion about that. And now I'm asking about Nicholas and Elizabeth. Well, I think that, yeah, I was going to try to tie this question to the previous question a little bit, too, in terms of thinking about ways that we can try to enhance public knowledge as a way to sort of mitigate some of these issues. One way is certainly by thinking about, if not getting informed consent upfront, certainly informing users still after the fact. And so doing things like sharing research outputs with our participants is a really good way of trying to let them know that this is happening, to let them see the outputs of this work and to try to build that sort of that trust with that participant community. I mean, I think certainly in terms of questions about there's an event and you have to respond to it quickly. And it's really important to maybe start the data gathering before you've necessarily done the full compliance side of the ethics portions. I mean, I think that, obviously, if there's a real experience to data collection and there's sort of an immediate, tangible impact and severity of it, I think that it's real important to start these processes and be in coordination with the IRB or the ethics review board or whatever it is as soon as humanly possible and let them know. I mean, there are certainly ways and IRBs have encountered this before of where you started the data collection and you're contacting them to get approval. Starting data collection, I should say public data and you're contacting them in parallel to get the process rolling. I think that's really important to make sure that if there is an experience to the data collection that it happens in tandem with the compliance side of things. And I want to just have a different take on the question and this conversation in that at what better time is there than right now to be doing research, right? I mean, think about the discourse right now around every day we're hearing about clinical trials. We're hearing about, you know, phase one trials, phase two when we're talking about vaccines and the development and so it seems to me that this year has almost been a crash course in public health education, right? For the whole world and it shows both, you know, some of the great things about our public health systems and of course it's shown where things are really, really terrible for many, many communities, many individuals where the public health systems have truly fallen. And so I think as researchers it's almost like, you know, this is our heyday, right? Like we have an opportunity to be talking about our ethics, our research. I mean, you know, again I can't think of a better time and either in the context of internet researcher or outside of it, I do think research has become, you know, very common. We're all talking about it right now and so perhaps it's the perfect time, you know, in the perfect storm we take opportunity, right? If I could, if we have about 10 minutes or so left, I had one question that I wanted to hear from all three of you and it got me thinking, Nick, when you showed the data about people's comfort levels when their tweets were analyzed by a computer, you know, system analytics versus by a human, that they were more comfortable with that, right? But then I go back to Mika Laj and what you showed us, right, where the machine learning was oftentimes so wrong. Okay, so there's those two pieces but then I also want to tie it back to you, Jan, where, okay, if we as researchers have some kind of ethical responsibility to intervene, perhaps in the case of depression or in the case of mental illness or in any case, in the case of any vulnerability, how do we tie all those three pieces together then? I don't have the answer and as soon as I saw that data it was like, uh-oh, where do we go with this? Yeah, well, so I think that's a good grant right there. I don't have an answer to the question. I mean, I think that it's important to understand why people feel more comfortable with having a machine analyze their content rather than a human and part of it is because of the fear of human judgment. One of the things, so one of the things we did in our survey is that we've left the opportunity available for people to include additional comments about the questions and one of the things that we got was very active mistrust of researchers, the belief that researchers were politically biased or that researchers are only out there for their own gain and so there was this idea that certainly seems to be tacit in the data that despite people who might have the STS background, who might have the knowledge that algorithms can be biased, that there is a more public belief that these things are neutral and that's something that we're going to have to sort of think through and deal with. And it's funny, again, to go to Michael, when I hear you say human judgment, I think of implicit bias that are then embedded in our systems and in our tools. Yeah, I think there are two opposing forces here at play, the one that Nick mentioned, the fear of human judgment and the other one which is basically giving or assigning agency to machines where there is no agency at all. For instance, if you give a messenger the access to your microphone, the messenger will eavesdrop on your conversation. So if on the phone you mention a certain brand, say I'm thinking about buying this and that, you're more likely than not to see the advert for that particular brand in your Facebook feeds two days in the future. But people will think someone is listening to my conversations, they have heard the brand, so that's why I'm saying this. Of course, nobody listens to that, even the American, even the Secret Service, which eavesdrops on every single conversation in the world, they also have just certain words or a combination of words that they listen to. But people somehow mix those two modalities. The idea of a machine which is impartial, which does not judge, which just something goes in and there is some rumbling and the number comes out. Basically, that's the idea, the public idea of a machine analyzing the data versus the researcher looking at the data and somehow judging me for my character, for what I've said, what I'm searching for. Before my death, please delete my browsing history, right? This is the most important thing you should ask, you should write down in your will, right? So nobody sees that. So, yeah, people are just very, very confused about what is being done to their data, who is doing this, what incentives there are at play, where the money is and what their input is. I would say I'd rather approach it from the economical point of view, trying to educate people about the economical play behind all of that because it's not nefarious. There are huge companies that just want to sell you more of their crap. That's what it is. So people should understand that and should understand that they are, if they're not paying their money, they're paying with their clicks, views, eyeballs, minutes of attention, which is the greatest price to pay given the finite amount of time that we have here on earth. So I would rather go into economy and trying to educate people on that level rather than trying to open the research and saying, yeah, please understand, this is a madness and this is how we do it and that's why we do it might be, I don't know if it's a better way, but I have a hunch that it will be more efficient. Yes, I was also thinking about the reason why people are so suspicious about researchers and I agree that they don't want to that they don't want to be judged, but when we are talking about also this obligations of researchers to act, we are somehow coming back to this double role of, for instance, a researcher who is also a physician and in the context of internet research, it does not happen. So researchers is just a researcher and the motive of research is quite mysterious for people. Why social researchers do research? Why computer researchers do research? We know why doctors perform research because and we want them to conduct as many research trials and observations as possible because we have hope also in medicine and that's why people also may think that they would like to contribute to biomedical research, but all this sphere of computer research and social research may seem to be very suspicious and somehow it conflates with this commercial aspect and that Miko, I describe that the whole, let's say, internet industry want to play our desires and want to make and want to press as much money as possible from our pocket. So on the other hand, I agree with Miko in this respect that education should be also about the business side, but I still think that this idea that also researchers share their intentions and also I like very much this idea to document ethical deliberation and ethical process because I believe that sometimes even when we make mistakes, but we make certain mistakes in a good faith, this somehow may not excuse us totally, but at least shows that we that we try our best to understand the problem and at least that we had a good good intentions. So this is probably not a very utilitarian approach, but I think that still it is important because also the procedure that we take in order to come to a certain conclusion seems to be important from the ethical point of view. You were muted. You would think after all these months, right? We're just about at time. All of our emails are available on the project webpage. If any of the questions didn't get answered, if you have further follow-up, feel free to reach out to any of us and I'll turn it back to you, Jan. Thank you. Now you're muted. Excuse me. The seminar came to its end, so I would like to thank, first of all, our participants. Thank you for all your questions and comments and for being with us. Then, of course, I would like to thank our great guests, Elizabeth and Nikolas and of course, Nikolai. I also want to thank Agnieszka Lempard, our administrative manager. She did a fantastic job organizing and advertising this event and without her, this event simply wouldn't happen. So thank you all. Thank you and goodbye.