When I was a teenager, my parents often asked me to come along to the store to help carry groceries. One day, as I was waiting patiently at the check-out, my mother reached for her brand new customer loyalty card. Out of curiosity, I asked the cashier what information they record. He replied that it helps them keep track of what we’re buying so that they can make tailored product recommendations. None of us knew about this. I wondered whether mining through millions of customer purchases could reveal hidden consumer preferences and it wasn’t long before the implications dawned on me: are they mailing us targeted ads?
This was almost two decades ago. I suppose the question most of us are worried about today is not all that different: how effective are micro-targeted messages? Can psychological “big data” be leveraged to make you buy products? Or, even more concerning, can such techniques be weaponized to influence the course of history, such as the outcomes of elections? On one hand, we’re faced with daily news from insiders attesting to the danger and effectiveness of micro-targeted messages based on unique “psychographic” profiles of millions of registered voters. On the other hand, academic writers, such as Brendan Nyhan, warn that the political power of targeted online ads and Russian bots are widely overblown.
In an attempt to take stock of what psychological science has to say about this, I think it is key to disentangle two prominent misunderstandings that cloud this debate.
Classic prediction models that only contain socio-demographic data (e.g. a person’s age), aren’t very informative on their own in predicting behavior. However, piecing together various bits of demographic, behavioral, and psychological data from people, such as pages you’ve liked on Facebook, results from a personality quiz you may have taken, as well as your profile photo (which reveals information about your gender and ethnicity) can improve data quality. For example, in a prominent studywith 58,000 volunteers, a Stanford researcher found that a model using Facebook likes (170 likes on average), predicted a whole range of factors, such as your gender, political affiliation, and sexual orientation with impressive accuracy.
In a follow-up study, researchers showed that such digital footprints can in fact be leveraged for mass persuasion. Across three studies with over 3.5 million people, they found that psychologically tailored advertising, i.e. matching the content of a persuasive message to an individuals’ broad psychographic profile, resulted in 40% more clicks and in 50% more online purchases than mismatched or unpersonalized messages. This is not entirely new to psychologists: we have long known that tailored communications are more persuasive than a one-size-fits all approach. Yet, the effectiveness of large-scale digital persuasion can vary greatly and is sensitive to context. After all, online shopping is not the same thing as voting!
So do we know whether targeted fake news helped swing the election to Donald Trump?
Political commentators are skeptical and for good reason: compared to a new shampoo, changing people’s minds on political issues is much harder and many academic studies on political persuasion show small effects. One of the first studies on fake news exposure combined a fake news database of 156 articles with a national survey of Americans, and estimated that the average adult was exposed to just one or a few fake news articles before the election. Moreover, the researchers argue that exposure would only have changed vote shares in the order of hundredths of a percentage point. Yet, rather than digital footprints, the authors mostly relied on self-reported persuasion and recall of 15 selected fake news articles.
In contrast, other research combing national survey data with individual browser histories estimates that about 25% of American adults (65 million) visited a fake news site in the final weeks of the election. The authors report that most of the fake news consumption was Pro-Trump, however, and heavily concentrated among a small ideological subgroup.
Interestingly, a recent study presented 585 former Barack Obama voters with one of three popular fake news stories (e.g. that Hillary Clinton was in poor health and approved weapon sales to Jihadists). The authors found that, controlling for other factors, such as whether respondents liked or disliked Clinton and Trump, former Obama voters who believed one or more of the fake news articles were 3.9 times more likely to defect from the Democratic ticket in 2016, including abstention. Thus, rather than focusing on just voter persuasion, this correlational evidence hints at the possibility that fake news might also lead to voter suppression. This makes sense in that the purpose of fake news is often not to convince people of “alternative facts,” but rather to sow doubt and to disengage people politically, which can undermine the democratic process, especially when society’s future hinges on small differences in voting preferences.
In fact, the second common misunderstanding revolves around the impact of “small” effects: small effects can have big consequences. For example, in a 61-million-person experiment published in Nature, researchers show that political mobilization messages delivered to Facebook users directly impacted the voting behavior of millions of people. Importantly, the effect of social transmission was greater than the direct effect of the messages themselves. Notably, the voter persuasion rate in that study, was around 0.39%, which seems really small, but it actually translates into 282,000 extra votes cast. If you think about major elections, such as Brexit (51.9% vs. 48.1%) or the fact that Hillary ultimately lost the election by about 77,000 votes, contextually, such small effects suddenly matter a great deal.
In short, it is important to remember that psychological weapons of mass persuasion do not need to be based on highly accurate models, nor do they require huge effects across the population in order to have the ability to undermine the democratic process. In addition, we are only seeing a fraction of the data, which means that scientific research may well be underestimating the influence of these tools. For example, most academic studies use self-reported survey experiments, which do not always accurately simulate the true social dynamics in which online news consumption takes place. Even when Facebook downplayed the importance of the echo chamber effect in their own Science study, the data was based on a tiny snapshot of users (i.e. those who declared their political ideology or about 4% of the total Facebook population). Furthermore, predictive analytics companies do not go through ethical review boards or run highly controlled studies using one or two messages at a time. Instead, they spend millions on testing thirty to forty thousand messages a day across many different audiences, fine-tuning their algorithms, refining their messages, and so on.
Thus, given the lack of transparency, the privatized nature of these models, and commercial interests to over-claim or downplay their effectiveness, we must remain cautious in our conclusions. The rise of Big Data offers many potential benefits for society and my colleagues and I have tried help establish ethical guidelines for the use of Big Data in behavioral science as well as help inoculate and empower people to resistmass psychological persuasion. But if anything is clear, it’s the fact that we are constantly being micro-targeted based on our digital footprints, from book recommendations to song choices to what candidate you’re going to vote for. For better or worse, we are now all unwitting participants in what is likely going to be the world’s largest behavioral science experiment.
Source: Scientific American
