Abstract
In this dialogue, Phillip Howard introduces “computational propaganda” as an emerging communication tool in political communication and a perspective for investigating misinformation and disinformation. By articulating the concepts, patterns, and mechanisms of computational propaganda, Howard proposes a socio-technical framework for studying computational propaganda. He calls for mixed methods to undertake computational research alongside qualitative investigation, thus addressing the computational as well as the political. Howard emphasizes the battle against algorithm bias, manipulation, and misinformation, and he advocates building an International Panel on the Information Environment (IPIE), an international scientific collaboration, to respond to the challenges. In addition, Howard offers advice on further research in computational propaganda.
Keywords
In recent years, “computational propaganda” has become an emerging communication tool in political communication, as well as a perspective for studying digital misinformation and manipulation. What does this concept mean?
Computational propaganda is the use of algorithms to distribute information that has been shaped in the service of ideology. As a communicative practice, computational propaganda involves algorithms, automation, and human curation to distribute content. The algorithms function with large computational systems, mobile phone networks, or social media platforms (Woolley & Howard, 2016a). And the problematic contents are those written to serve a specific purpose: either a political agenda or someone’s interests. Sometimes, it is hard to understand the political purpose; sometimes, it’s straightforward—you can tell who’s behind it and who’s paying for it.
Although computational propaganda seems to have a short history, it has a certain continuity with the existing propaganda institution in a given society. In other words, the term “propaganda” is an apparent grand ancestor of this modern concept of computational propaganda. How does computational propaganda build its own institutional infrastructure, which might be similar to (or different from) a traditional propaganda infrastructure?
Nesting upon the tradition of propaganda, computational propaganda employs similar psychological mechanisms. It appeals to the user’s emotions and biases in order to circumvent rational thinking and promote specific ideas outlined by the developers of such a campaign. Thus, similar to the conventional propaganda practices of the past, computational propaganda creates and directs public opinion as a gradual process conducted by ad-hoc methods.
However, computational propaganda expands the traditional misinformation toolkit and substantially differs from traditional propaganda in terms of the methods and scale of the manipulation campaigns. One advantage of computational propaganda infrastructure is its ability to change the message and then personalize that message for a particular audience. So, the audience for computational propaganda is not the broad public: it’s not the entire city of Hong Kong or the entire country of Canada. It is a fairly small segment. The art of computational propaganda is identifying the few thousand of each type and getting them the right message.
One of the lessons that we have learned recently about the news industry in Europe is that several news outlets are purposely designing content for European policymakers in Brussels. The average reader of these websites is not the European public; it is not the people of Belgium; it targets a relatively narrow audience. Another example would be the ongoing war in Ukraine: publishing news stories that present different perspectives on Russia’s invasion of Ukraine and targeting an audience that you know will already have an opinion, respectively, that’s part of the contemporary mechanism of computational propaganda.
What are the typical formats of computational propaganda?
Well, there is much nuance these days. Sometimes the story is entirely fake, completely ridiculous, and nothing is accurate. Sometimes the photo is accurate, but the text isn’t; using the BBC colors or the New York Times font to make it look like a news article. Sometimes it’s very subtle, most of the facts are correct, but there’s an overemphasis on the negative or a particular perspective. We might call it just biased journalism, or we might call it “mal-information.” In recent years, especially during the pandemic, the types of computational propaganda have diversified.
An emerging field calls for a theoretical lens for investigation. In your edited volume, Computational Propaganda: Political Parties, Politicians and Political Manipulation on Social Media, you frame the computational propaganda from both technology and social perspectives. What does this socio-technical framework imply?
The synergy of big data, algorithms, and autonomous agents or so-called “social bots” assemble computational propaganda as a technical term (Woolley & Howard, 2016b). Hence, technical components such as automation, scalability, and anonymity are vital in distinguishing computational propaganda from existing practices.
However, it is a dangerous position to understand computation propaganda only from a technical perspective. The process of computational propaganda is more than a set of variables, models, code, and algorithms. This powerful new communication tool is produced within the systems of power and knowledge (Bolsover & Howard, 2017). Thus, if we aim to examine and respond to the impacts of computational propaganda on our political system and public life, there is no other way but to undertake computational research alongside qualitative investigation, thus addressing the computational as well as the political.
Such a socio-technical framework has implied and invited mixed methods when investigating computational propaganda. In this edited volume, researchers have adopted various methods, including computational methods, ethnography, surveys, political opinion polling, in-depth interviews, and comparative policy analysis.
Yes, our research team for this multi-year study of computational propaganda worldwide used both qualitative and quantitative methods. The team involved “12 researchers across nine countries who, together, interviewed 65 experts, analyzed tens of millions of posts on seven different social media platforms during scores of elections, political crises, and national security incidents.” (Woolley & Howard, 2018, p. 11) Researchers adopted different methods depending on the context of each case.
Through international comparisons, do you identify the general patterns of how social bots function across different societies?
Well, I think there are only a few broad generalizations that are safe for countries in North America, Latin America, and perhaps, Africa. Two years ago, we did a study with Gallup, a polling company, to ask people around the world—there were 140 countries in this study—what fears they faced in using the Internet. We asked a whole set of fear-related questions such as sexual harassment online, gender harassment online, credit card fraud (whether people were afraid of using a credit card), exposure to misinformation, and the fear that technology would take our jobs. The number one consistent concern was the fear of being misled: 52 to 53 percent of the respondents and 150,000 respondents said that they fear being misled by something they found online. The sources of misinformation these days tend to be regular political parties or politicians running for office because they are the ones who have a campaign budget, and they will spend money to get elected.
But there was a lot of regional variation. In Latin America, most respondents (at least the women) feared sexual harassment online through digital technologies. In Southeast Asia, the fear of being misled was not so much of a concern. Thus, on the one hand, there are a few global tendencies. On the other hand, there is cultural variation in what people consider the most valuable in technology and the most significant risk of the technology network.
Across different societies and cultures, there are usually intertwined mechanisms for computational propaganda.
Indeed. It’s always good to talk about examples. In the last few years, we have spent time studying Covid misinformation (Brennen et al., 2020). Some incentives seem to be mostly about making money: people will spread misinformation and try to sell you a baseball hat or try to sell you a T-shirt. Sometimes they want to sell fake cures or get you to protest lockdown measures; we certainly saw this in the United States. The motivation for generating the propaganda wasn’t so much about speaking truth to power or questioning the research, but trying to make some money. For the example of Covid misinformation, that’s the primary mode we found.
In this case, it is an intertwined political agenda and a commercial one to produce the misinformation that brought the infodemic into the pandemic. Many researchers have engaged in exploring a “cure” for misinformation. From a perspective of computational propaganda, what are the solutions we’re proposing here? Do we even have a solution?
Well, I think a part of the solution must involve some regulation. We have passed the point of industry selling self-regulation. And many of the technology firms in Silicon Valley and internationally now say they would like some policy guidelines to know what to design for. One of the tricks is to develop regulations that don’t stifle political speech but make it possible to prevent the average social media user from getting the worst of the content. So, in statistical language, maybe we would say: “I want to decrease the probability that large numbers of people will encounter junk when they’re doing their social life online, and increase the possibility that they’ll encounter good ideas and accurate information.” This is not just so much about computational propaganda but also algorithm bias and algorithmic manipulation.
Are you concerned the call for algorithm regulation might become an abuse of the government’s administrative power?
It may go country-by-country. Most democratic countries already have rules on the books about how an election should be run. But not all countries give their regulators the resources and the personnel. Sometimes they can fine a political party, but it is a very small fine, so it’s easier to try and do the trick and face the fine later. So, in practice, I think this has to be about equipping the public agencies we already have with the resources to do their missions.
However, situations in the crisis might be different. When the government needs to get out a public health message or needs to get out a message about a complex natural disaster, any interference with that message might cost lives. Those are sensitive moments when we need the platforms to behave well, collaborate, and share data with researchers to help save lives.
What kind of collaboration do you suggest?
At the nation-state level, I think global cooperation is possible, and there are a couple of institutions already. China has been the chair of the International Telecommunication Union for the last two years, and several important conversations have happened. There is the Internet Governance Forum, another body—sometimes it’s a bit slow, but it does work towards consensus building.
I think there are also ways that we, researchers, can help, so those of us who are interested in, who are able to do the data science, can find ways to link up, to help each other, and make a case for more data, demonstrate when the algorithms are having bad outcomes or unintended outcomes.
One of the things I’m getting excited about now is the possibility of having scientific communities in many countries collaborate to identify the kinds of data that they need to be able to answer important social and public policy questions, simplify that list and then present it to the firms, so they are able to say: “We need this data to solve these problems. Please give.” And if the firms don’t want to offer the data or they provide the data, but it’s low-quality data or not quite what we expected, we need to be able to draw public attention to this as a problem. So, I do think there are very important researchers in the UK, Canada, China, and around the world who could do more good work and could advise their governments if they had access to the data about social problems. Right now, the best data on public problems is not in the public realm. It doesn’t sit in our lives—it’s not in your or my library—we don’t have access to it. It’s in Silicon Valley. Until we can crack open some data and some data access issues, I think we will be stuck on many policy questions.
What types of data are the top three on your wish list?
At the moment, we get access to so little. For example, if a researcher wants to study messaging about Covid and if a public agency wants to encourage people to wear masks for a brief moment, it’s crucial to have a full cycle of data about the messaging around “please wear more masks and here’s why” to see how the messages get changed, shared, engaged with, repackaged and repurposed. With such data, researchers can understand where the messages that discourage masks or discourage people from collaborating with health rules come from, and why they are appealing. That kind of data might be qualitative, which might come from fieldwork more than quantitative data.
So, the data ecosystem becomes crucial not only for computational propaganda but also for computational social science in general. Otherwise, we can only compute what can be easily measured, which might produce incomplete, and even biased, pictures of the social problems.
At the moment, I’m spending a lot of energy working on how to network, bring scientists together to study misinformation, and make some demands of technology firms to help fix some of the big problems such as climate change, human trafficking, and global health. Economic development is another domain in which algorithmic interference can shape businesses and the whole market. I’m trying to get our research community together to have the researchers from China meet with the researchers from Canada and the researchers from Africa. That should be the next step we must take with our research communities.
What are the challenges of studying computational propaganda?
There are several challenges. One is about methodology. In the first wave of research on computational propaganda, most of us were looking at the text: simple text, SMS messages, or highly automated activity on social media platforms. Now it’s much more about visual content, visual cues, and that’s much harder to study. So many of us are good with the text but challenged by the images. I think this is one area where social scientists can learn from communication research because scholars in communication and journalism have been studying video images and still images for a long time with coding schemes. You can make generalizations from these kinds of big studies.
At the moment, we don’t have an image archive. In many countries, one of the most basic resources civil society groups and academics ask for is an ad library—a collection of the ads that go out—political ads or others. And not just some ads, but all the ads. That’s the only way to understand an information ecosystem comprehensively.
Another challenge of computational propaganda is data. The communication content is often “A/B tested.” The message is tested on a small population. You see the click-through rate, distribute the most successful version, and then do an A/B test again so that the image or text changes rapidly over time. We, as academics, don’t get to see the content that was produced on Day 1, Day 2, Day 3, and Day 4—as it changes and evolves. As researchers, we only tend to get the outcome or a selection of artifacts.
When we get that data, it often does not include network information or engagement statistics. The data we get are not always meaningful when we get them from the firm. It’s a quandary. People who like to do extensive and statistical analysis find the data doesn’t always meet their expectations for high-quality data. The people who do qualitative work or critical work can get access to the fields of the companies that make samples of content, but not large pieces. And they can’t always generalize from the things they see.
The other challenge is that most of us work on computational propaganda in English. Still, we know the data also exists in Arabic, Spanish, and many languages worldwide. And the technology firms don’t share that data either. I think the ability to work across languages needs to be the future of this line of research. Understanding what’s happening in multiple languages at once is a big challenge for us.
If you would advise newcomers who want to study computational propaganda, what advice do you have for them?
Well, first, I would say very broadly that it’s worth focusing on social problems, and we’ll take a social problems frame and try to help solve social problems. Second, try to be multimethod, open to critical theory, and large-scale collaborative work. You can work in your political culture and learn something about your neighbors. So, understanding the cultures around you is going to be critical. Third, I think that the operational challenges we now face are data with multimodality; it is worth spending time working on the problem of understanding the effects of visual context. And then I would also suggest studying the platforms that don’t get studied very much. We all spend a lot of time on Twitter and Facebook. There are so many other platforms that people use, the kids use, and I don’t even use them. Trying to understand those other platforms in the comparative context is very important.