Facebook’s anti-fake news measures are being tested during the Covid-19 outbreak. Social media has become a hotbed of misinformation as the virus spreads worldwide, and since this misinformation relates to public health and has the potential to influence a person’s decision to seek treatment, social platforms have been forced to alter their approaches to prevent the spread from worsening.
Always the distributor and never the editor, Facebook has rarely taken to removing content unless it is in clear violation of its rules – an approach that has garnered much criticism, especially when it comes to political advertising. But during the novel coronavirus, removal has become commonplace. Facebook only removes misinformation "that may contribute to physical harm"; other types of misinformation have their distribution reduced. It has been focusing on claims that have been debunked by the World Health Organization or other credible health experts, and are most likely to result in someone getting sick or not seeking treatment.
Facebook has also introduced new tools during the Covid-19 outbreak, including a pop-up that links to credible health information that will be surfaced when people search for information related to the virus on the platform or tap a related hashtag on Instagram. It was launched in Hong Kong, Japan, Macau, Malaysia, the Philippines, Singapore, Taiwan, Thailand and Vietnam.
The company has also been sharing aggregated and anonymised mobility data and high-resolution population density maps to researchers at Harvard University’s School of Public Health and National Tsing Hua University in Taiwan to help inform their forecasting models for the spread of the virus.
But Facebook is mainly reliant on its fact-checking program to stymie the spread of fake news related to Covid-19. In Asia-Pacific, Facebook has 27 fact-checking partners across 11 countries and territories. AFP Fact Check is a partner in 10 of the 11 territories (excluding Taiwan) and is the sole partner in five territories: Hong Kong, Malaysia, Pakistan, Singapore and Thailand. This is slightly more than the 26 fact-check partners Facebook has in Europe across 16 countries. It has seven partners in the US.
So how effective is Facebook’s fact-checking program in identifying and reducing misinformation? Campaign Asia-Pacific has interviewed AFP Fact Check and analysed a number of coronavirus-related Facebook groups to paint a clearer picture of its process and its flaws.
The fact-check partner view
AFP Fact Check was established in 2018 in response to a "growing tide of online disinformation", specifically on social media, according to Rachel Blundy, Hong Kong editor at AFP Fact Check. While it has worked through various periods of misinformation spikes in the past – such as during the Hong Kong protests – the Covid-19 outbreak is likely the biggest story it has worked on so far.
The service has seen a "wave of misinformation" in Asia about Covid-19 since mid-January, Blundy says. Misinformation has ranged from prevention to cures to xenophobic claims.
"Initially, we saw a lot of misinformation about the origins of the virus and how it was affecting people in the Chinese city of Wuhan," Blundy explains. "Then, as the virus started to spread to other countries around the world, we have seen misleading social media posts about how people can prevent themselves from becoming infected, as well as alternative 'cures'.
"We've seen images and videos being taken out of context or misrepresented throughout the outbreak. A lot of posts have been xenophobic in tone, suggesting the virus has a specific connection to people of Chinese ethnicity, which it obviously doesn't."
Coping with the surge in demand "has certainly been a challenge", Blundy says. For the most part, AFP has only one fact-check reporter in each of its 10 Asia-Pacific territories, as well as a team of five editors in Hong Kong. During the outbreak, it has been leaning on support from AFP’s 20 bureaus outside the region, as well as staff from the wider AFP network, which includes more than 1,700 journalists in 201 bureaus across 151 countries.
"That’s allowed us to maintain a steady flow of reports on misleading posts on the virus from multiple datelines around the world," Blundy adds.
Since mid-January, AFP’s Asia-Pacific team has published about 60 fact checks on coronavirus-related content to date, while bureaus from the rest of the world have fact-checked about 100 additional pieces of content. Blundy estimates that about half of the claims AFP has fact-checked on coronavirus has included a misleading Facebook post.
But AFP reporters only find "about 30-40%" of their stories from Facebook’s dedicated fact-check feed, Blundy continues. Reporters have to balance their time across all platforms and information sources. When they are not looking through Facebook’s fact-check feed, they sift through other platforms such as Twitter, YouTube and Weibo, web pages and articles. Using keyword searches is an effective tool, as it allows the reporters to source claims that have appeared in multiple social platforms, such as a video claiming to show a murder of crows in Wuhan that had been shared on Facebook, Instagram, Twitter and YouTube and which has been debunked by AFP.
Reporters also curate lists on Facebook-owned social-monitoring platform Crowd Tangle. Usually used by publishers to keep an eye on trending news, AFP uses it to track "trending disinformation", according to Blundy. It allows fact-checkers to keep tabs on repeat offenders who share the same misinformation across multiple pages and groups.
How Facebook’s fact-check system works
Facebook says the system it has to identify fake news is a "hybrid between people and technology". In the first instance, it relies on its community of users to report content they see as "false news". Content that can be fact-checked includes ads (except political ones), articles, photos or videos on Facebook and Instagram, as long as the content is public.
A machine-learning algorithm then sifts through user-flagged posts – and the wider Facebook ecosystem – to scan for links to disreputable websites in order to prioritise posts that are to be sent to third-party fact-checkers.
Content that is flagged for review is collated in a "claim check feed" that fact-checkers can toggle by geography and language. The feed is refreshed on a weekly basis. But, beyond that, there isn’t much organisation to the feed. Content appears in no particular order and makes no suggestions on which posts should be prioritised – fact-checkers are left to decide for themselves. This respects the journalistic integrity of the fact-checkers but also means that often the same claim/post is fact-checked more than once by different partners or in different languages, wasting resource in the process.
The algorithm that populates the feed is "not perfect", according to Blundy. She says "quite a lot" of the feed is populated by the wrong kind of content: content that is incorrectly tagged or that can’t be fact-checked because it is conspiracy or opinion, along with legitimate news stories from respected publishers.
"Sometimes the moderation system is pulling in violent content or sexual content that people don’t want on Facebook. We also see conspiracy theories start to emerge relatively early on, because there is so much panic and anxiety, but we can’t fact-check a lot of this because there is no evidence to support it and no way to verify it," Blundy continues. "And a whole range of media outlets have popped up in there. The understanding of what is fact-checkable is not quite there yet."
This is problematic, because while Facebook waits for fact-checkers to work through the feed, posts that have been flagged as potentially false – either by users or algorithms – have their distribution reduced. This could mean suppressing news stories containing vital information about Covid-19, for example.
Once a fact-checker has reviewed a post in the feed, they can give it one of nine ratings: false, partly false, true, false headline, not eligible, satire, opinion, prank generator and not rated. Any content that has been flagged under one of three possible fake news ratings is demoted in the news feed to reduce distribution and users are notified of the rating when they click to share it. Repeat offenders are commonplace on Facebook and the platform will sometimes take down pages that have been flagged multiple times – or at least remove their ability to monetise.
The fact-checker ratings help to train the machine-learning algorithm to spot potentially false content to reduce the reliance on user flagging. The machine-learning model can also identify duplicates of debunked stories. Fact-checkers are asked to focus on "the worst of the worst" – that is, clear misinformation and fake news intended to harm and mislead. Facebook has four criteria it asks fact-checkers to consider when prioritising what content to check: verifiability (claims based on facts rather than opinion), importance, relevance (to news or current events) and virality.
Then on to the fact-checking, which is a cumbersome process. Algorithms and digital tools may help surface potentially fake news, but the actual process is manual and resource-heavy. AFP reporters scrape videos and images for metadata, scan videos frame by frame for insignia or dialects that may give a location away, conduct reverse image searches and combine this with regular journalistic practices such as research, contacting the original sources and obtaining official statements/police reports.
So what's the solution?
It is clear that Facebook’s fake-news fight has significant flaws, although the proportion of content it catches and removes has likely increased (it has not provided official data). User flags are proving problematic due to the newness of the "fake news" phenomenon, with many not understanding what it is or abusing the tag to discredit publishers they don’t like. It is why both platforms such as Facebook and news organisations such as AFP are focusing on improving the general public’s news literacy.
"Misinformation cannot be eradicated, but media literacy can be boosted to help people avoid being misled. The work being done now will leave the next generation much better equipped to identify misinformation/disinformation online," Blundy says.
Improving media literacy, especially in developing countries, will help to "cultivate a free and fair media environment" more than legislation, she believes.
"Anti-fake news laws can have an impact on the amount of misinformation online in a particular country, but they can also be used to stifle dissenting voices. Some countries have acknowledged that these types of laws don't work particularly well in practice," she adds.
While Facebook continually announces fresh investments in its anti-fake news measures, Blundy says her job over the past year "hasn’t gotten any easier". A large part of this comes down to resource: journalists are grossly outnumbered by peddlers of fake news.
Media literacy is a long-term solution, but hiring more fact-checkers is more pressing. Facebook would not disclose how much it pays its fact-check partners, but a recent investigation by newsletter Popular Information found that it paid a top US fact-check partner just $359,000 in 2019.