Nahema Marchal is a doctoral candidate at the Oxford Internet Institute, University of Oxford, and a researcher for the Computational Propaganda Project. Her research examines the relationship between social media and polarization and the manipulation of digital platforms in the context of mis- and disinformation campaigns. She is also an experienced media spokesperson and regularly provides insights to outlets including the New York Times, Politico, the Financial Times and the BBC.
Can you tell us a couple of things about the research on Covid-19 disinformation that you have been conducting at the Oxford Internet Institute? Is OII conducting any longitudinal research in this context?
Since the beginning of the Covid-19 outbreak, our team of researchers at the Oxford Internet Institute has led a multi-pronged effort to map out the spread of coronavirus information across multiple platforms. As part of that effort, we have been publishing weekly briefings on real-time news trends, and led several data-focused studies: one looking at YouTube video content, and others focused on state-based media outlets’ coverage of the crisis. Right now, we are finalising the third in our series: a memo on the visibility and discoverability of junk health sites in searches.
Why did you choose to focus on YouTube for that recent data memo? Can you talk a bit about its findings? Also, you mentioned one of the constraints of that research was the issue of personalisation. Do you think increasing personalisation will hamper effective online media research going forward?
YouTube is a key player in the online information ecosystem — and in recent years, it has also become a major source of information about health, science and tech, especially for young people. The average UK adult, for example, spends at least 30 minutes watching videos on YouTube every day. With this in mind, we decided to investigate what type and quality of information users would likely stumble upon when making simple searches for information about the virus.
Specifically, we analysed video results returned from around four search queries that were popular in the UK around early March — right before the lockdown measures were applied nationwide.
We developed a rigorous, multi-step classification process, almost like a “traffic light” system, focused on the type of the channel that was sharing videos and how politicized, confessional and factually accurate the information it relayed was.
There were a number of striking findings. First, we were reassured to find that the bulk of top video results came from professional news channels, who in most cases shared factual reporting and official recommendations from the World Health Organization (WHO) or other health officials. We were surprised to find however, that none of this information came directly from health agencies themselves — despite them having a presence on the site.
Reassuringly, few of the top videos returned for our search queries contained misleading or junk health information. However, this content is far more likely to encourage engagement in the form of comments from those who choose to view it.
We also found a fair number of politicized videos, some of them comedy shows, and other independent investigations into purported “cover-ups” about the magnitude and lethality of the Covid-19 epidemic. Many videos also engaged critically with the Chinese government’s policy on wildlife trade and sensationalized the virus’ origins.
You also mentioned the proprietary nature of YouTube’s search and recommendation algorithm being a constraint in your research. Isn’t the algorithmic nudging that YouTube employs, feeding off your previous behaviour (something that is not that pronounced in incognito mode) the main problem with disinformation? Do you think policymakers will need access to that eventually, so researchers can inform their work?
Not so much a constraint as much as a limitation. Personalization is built into the very fabric of our experiences online, and there is not much researchers can do to account for that— or neutralize it.
There’s been a lot of research done on YouTube’s recommender systems in recent years, including several algorithmic audits. The reality, however, is that, these systems are extremely difficult to study (due to their proprietary nature and constant changes) and we don’t have a good enough handle on how they work at any given time. Crucially though, while it is true that recommendation algorithms draw, at least in part, on users’ preferences and past behaviours to push content in front of their eyes, this is far from the only path to misleading or radicalizing content.
YouTube is as much a community as it is a portal to information. What this means is that viewers often follow creators that they like and trust — from political commentators to religious leaders and self-declared health experts — with little regard to the neutrality or factual accuracy of their content. In fact, as media scholars like Rebecca Lewis and Dr Alice Marwick have shown, it’s quite the opposite. More often than not, a user might follow a specific influencer precisely because they position themselves as a pariah or reactionary, or because they share non-mainstream news. And it’s pretty easy to push deceptively subversive messages, without sharing outright lies, to dedicated audiences who take it at face value.
In other words, people who seek junk content, because they find it entertaining or are simply curious, will always find it, as long as they know how to look for it. The problem is that the content is there in the first place.
To what would you attribute the limited presence of content by public health organisations like the NHS or the WHO that you mentioned?
It’s difficult to attribute this to one specific cause. How prominently YouTube videos appear on the site depend on a number of things – both how many people engage with them and similar videos, whether channel owners promote them, how they make use of keywords — one, or any combination of these factors, could be to blame. But more broadly, in the age of star influencers and search-engine optimized content, where people seek both informative and entertaining content, it’s possible that traditional institutions are simply pushing content that’s too nuanced or not compelling enough for viewers to engage with.
You have also researched the social media accounts of state-backed media outlets from China, Russia, Turkey and Iran mentioning that some have distribution networks that can reach hundreds of millions of people. At the same time you have identified an effort to portray Western democracies as incompetent in dealing with the coronavirus. Can you talk a bit about these findings?
For this memo, our team examined how news outlets backed by the governments of China, Iran, Russia and Turkey covered the Covid-19 crisis over a two-week period — both what they were writing on, and how large their audience on social media was.
We found that, while these outlets are not as prolific as other household names, they have a substantial global audience, pushing content to French, German and Spanish readers for instance. And they can achieve as much as ten times the effective engagement on the articles they publish.
What’s interesting about these outlets is that they do not always push outright disinformation. Most of the content they produce is factual, but distorts political reality in a way that suits their agenda, subtly promoting the idea of a decline of Western democracies for instance, or by exacerbating already present social tensions and antagonisms.
Tech platforms have taken a series of measures to address Covid-19 and you did mention for example YouTube’s decision to demonetise but then monetise again coronavirus-related content. What do you think about the tech platforms’ policies in terms of speed, consistency, efficacy and transparency?
Digital platforms’ concerted efforts to tackle coronavirus misinformation, and the speed with which they reacted, is certainly praiseworthy. If anything, it shows that with enough urgency, will and concern for public safety, they can pivot their policies very quickly to address loopholes that could harm the public.
One of the downsides of these quick fixes and ad hoc rules, however, has to do with transparency and efficacy. First, we need to understand the potential ramifications and unintended consequences of any given policy change before implementing it, especially when these impact millions of users. Beyond that, it is important to put transparency and accountability mechanisms in place to evaluate how effective these changes are in curtailing the problem. There is also a question of consistency: how consistently are these policies being applied, across platforms and across geographies?