How social media filter bubbles and algorithms influence the election
With Facebook becoming a key electoral battleground, researchers are studying how automated accounts are used to alter political debate online
One of the most powerful players in the British election is also one of the most opaque. With just over two weeks to go until voters go to the polls, there are two things every election expert agrees on: what happens on social media, and Facebook in particular, will have an enormous effect on how the country votes; and no one has any clue how to measure what’s actually happening there.
“Many of us wish we could study Facebook,” said Prof Philip Howard, of the University of Oxford’s Internet Institute, “but we can’t, because they really don’t share anything.” Howard is leading a team of researchers studying “computational propaganda” at the university, attempting to shine a light on the ways automated accounts are used to alter debate online.
“I think that there have been several democratic exercises in the last year that have gone off the rails because of large amounts of misinformation in the public sphere,” Howard said. “Brexit and its outcome, and the Trump election and its outcome, are what I think of as ‘mistakes’, in that there were such significant amounts of misinformation out in the public sphere.
“Not all of that comes from automation. It also comes from the news culture, bubbles of education, and people’s ability to do critical thinking when they read the news. But the proximate cause of misinformation is Facebook serving junk news to large numbers of users.”
Emily Taylor, chief executive of Oxford Information Labs and editor of the Journal of Cyber Policy, agreed, calling Facebook’s effect on democratic society “insidious”. Taylor expressed similar reservations about fake news being spread on social media, (a term Howard eschews due to its political connotations, preferring to describe such sources as “false”, “junk” or simply “bad”), but she added there was a “deeper, scarier, more insidious problem: we now exist in these curated environments, where we never see anything outside our own bubble … and we don’t realise how curated they are.”
A 2015 study suggested that more than 60% of Facebook users are entirely unaware of any curation on Facebook at all, believing instead that every single story from their friends and followed pages appeared in their news feed.
In reality, the vast majority of content any given user subscribes to will never appear in front of them. Instead, Facebook shows an algorithmic selection, based on a number of factors: most importantly whether anyone has paid Facebook to promote the post, but also how you have interacted with similar posts in the past (by liking, commenting or sharing them) and how much other people have done the same.
It is that last point that has Taylor worried about automation on social media sites. Advertising is a black hole of its own, but at least it has to be vaguely open: all social media sites mark sponsored posts as such, and political parties are required to report advertising spend at a national and local level.
No such regulation applies to automation. “You see a post with 25,000 retweets or shares that comes into your timeline,” Taylor said, “and you don’t know how many of them are human.” She sees the automation as part of a broad spectrum of social media optimisation techniques, which parties use to ensure that their message rides the wave of the algorithmic curation on to as many timelines as possible. It is similar, though much younger and less documented, to search engine optimisation, the art of ensuring a particular web page shows up high on Google’s results pages.
Who should I vote for? The UK election manifestos compared
Academics such as Taylor and Howard are trying to study how such techniques are applied, and whether they really can swing elections. But their efforts are hurt by the fact that the largest social media network in the world – Facebook – is almost totally opaque to outsiders.
If Howard’s group were examining Facebook rather than Twitter, they “would only be able to crawl the public pages”, he said. That would miss the vast majority of activity that goes on on the social network, on private timelines, closed groups, and through the effect of the algorithmic curation on individual feeds. Even so, he says, those public pages can be relevant. “In some of our other countries studied, we think we’ve found fake Facebook groups. So there are fake users, but the way we think they were used – with Trump in particular – is that they were used, created, hired, rented, to join fake fan groups that were full of not-real people.
“Those fake groups may have eventually attracted real fans,” he said, who were emboldened to declare their support for the candidate by the artificially created perception of a swell in support for him. “There’s all these Trump fans in your neighbourhood, that you didn’t really know … so we think that’s the mechanism. And then we think some of those public pages got shut down, went private, or just because so full of real people that the fake problem went away. We don’t know, this is the theory.”
Analysis Facebook’s failure: did fake news and polarized politics get Trump elected?
The company is being accused of abdicating its responsibility to clamp down on fake news stories and counter the echo chamber that defined this election
Facebook does allow some researchers access to information that would answer Howard’s questions – it just employs them first. The company publishes a moderate stream of research carried out by its own data scientists, occasionally in conjunction with partner institutions. By and large, such research paints a rosy view of the organisation, though occasionally the company badly misjudges how a particular study will be received by the public.
In 2014, for instance, the social network published research showing that two years earlier it had deliberately increased the amount of “negative” content on the timelines of 150,000 people, to see if it would make them sad. The study into “emotional contagion” sparked outrage, and may have cooled Facebook’s views on publishing research full stop.
As a result of their lack of access to Facebook, Howard and his team have turned to Twitter. Even there, the company’s limits hit hard – they can see just 1% of posts on the site each day, meaning they have to carefully select what terms they monitor to avoid being too broad. In the US election, they hit the cap a few times, missing crucial hours of data as conversation hit a fever pitch.
Similar limitations exist throughout the study. The team had to use a broad definition of “automated posting” (they count any account that makes more than 50 tweets a day with political hashtags), because Twitter would not share its own definition. And they had to limit their examination of political postings to tweets that contain one of about 50 hashtags – such as #ge17 – not only to avoid hitting the 1% limit, but also to only scoop up tweets actively engaged in political debate.
The result, Howard said, was “vaguely analogous” to the conversation on Facebook: while it may not be the same, it is likely that debates that are most automated on Twitter are also most automated on Facebook, and in largely the same direction.
But the limitations have one huge advantage: unlike nearly every other academic in the world, and the vast majority of civil institutions responsible for regulating the fairness of elections, the Oxford Internet Institute aims to publish its findings before the election, updated on a regular basis. When I met Howard, along with two of his DPhil students and fellow researchers John Gallacher and Monica Kaminska, they were just beginning to make sense of the Twitter data. The graduate students face a heavy couple of weeks coding the data, manually marking web domains and links into categories like “news”, “hoax” or “video”, but the result should be a rare glimpse into a style of campaigning that many of its practitioners wish was hidden.
It is too early to say what results they will get, though Howard has pulled one finding from the preliminary data: judging by their own metrics, there do not seem to be significant amounts of bots posting Russian content, such as links to Sputnik or Russia Today.
Even there, though, he wishes for a small amount of extra cooperation. “We’ve stopped working with geolocation,” he said, referring to the process of trying to work out from where a particular tweet was sent, “but Twitter has the IP addresses of every user.” Sharing that, even in aggregate, anonymised form, could shine a tiny light on a side of democratic politics shrouded in darkness. These days, on the internet, no one knows you’re a bot.