April 29, 2014

The strange truth about fiction

By: Lada Adamic, Adrien Friggeri, Justin Cheng, Dean Eckles

Sharing rumors can be useful: they can be important sources of truthful information, especially in situations where authoritative sources of information are unavailable. Because rumor sharing is such an important activity in social networks, nearly every technological advancement that facilitates human communication, such as Facebook’s “Share” functionality, is ultimately used to relay rumors. In this note we examine a set of rumors propagating on Facebook either as text or images. One example of an instance of such a rumor is the following photo:

This photo has been posted and reshared hundreds of times in 2013, often accompanied by captions such as “War is a sad and tragic event for everyone except those who profit” and dating the bike to 1914. In reality, the bicycle is not nearly as old and has nothing to do with war: it has been identified as belonging to one Don Puz who abandoned it in the woods on Washington’s Vashon Island in 1954.

Intrigued by the persistence of such stories, we undertook a quantitative study of rumors circulating via Facebook and found the following:

  • Rumors propagate readily, with the same rumor often being uploaded multiple times, and some of those uploads being reshared tens or even hundreds of thousands of times.
  • Being made aware that a post is a known rumor prompts some individuals to retract their post, especially if the rumor is false; however, the propagation is largely unaffected.
  • There are instances of rumors evolving to make themselves more current, as well as counter-rumors spreading in the footsteps of a rumor.

The details can be found in our Rumor Cascades paper, to be presented at ICWSM’14 and we outline some of the results below.

Tracking rumors on Facebook requires two types of information: a corpus of known rumors, and a sample of reshare cascadescirculating on Facebook which can be matched to the corpus. The website Snopes.com has diligently documented thousands of rumors, and provides the starting point for our analysis. To match known rumors to this anonymized set of reshare cascades, we identify uploads and reshares that have been snoped — someone linked to a Snopes.com article in a comment. Those comments are posted by people to either warn their friends that something they posted is inaccurate or to the contrary, to validate that a rumor, though hard to believe, is in fact true.

We gathered 250K comments, posted during July and August 2013 on 17K individual cascades, containing 62 million shares. One large cascade shows a rumor diffusing as reshares prompt more reshares, forming long chains:

Reshares of a photo of a shopping receipt purporting in which someone buying sporting goods were paying an extra tax due to the Affordable Heathcare Law; this was due to a bug in software used by Cabela’s. One can observe long chains of reshares of the photo.The cascade above was frequently snoped, as can be seen from the lower-right branch with the reshares being snoped highlighted in red.

Although false rumors are predominant (62% of cascades have been tagged by Snopes as false), we observe that true rumors are more viral, in the sense that they result in larger cascades, achieving on average 163 shares per upload, whereas false rumors only have an average of 108 shares per upload.¹

So how do people react when their reshares are snoped? Especially if the rumor is false, there is a greater likelihood that they will delete the reshare (they are 4.4 times more likely to delete it). However, they are also more likely to retract the share even if the rumor is true or partly true, potentially because they realize that the story is old.

In general people do react to rumors with skepticism, though some true rumors evoke expressions of amazement and respect, as can be seen from the frequency of words that distinguish comments posted in response to rumor photos contrasted with comments on photos in general. The darker words are more likely to be associated with true stories, the lighter ones with false ones:

Some of the rumors circulating as photos actually originated years ago. For example this photo of the “money bags” meme was reshared over 100,000 times in July of 2013.

The meme had been circulating for years, and had been copied and pasted as a status update prior to the “Share” functionality. The meme purports that this year is special because some month has e.g. 5 Fridays, 5 Saturdays, and 5 Sundays. On average, this happens in one month each year, but the meme claims it happens only once every 823. In this case we see that the meme switches months/days to be accurate in the current year. Another striking feature is the flare-up of the rumors: they all but disappear, but can re-ignite secondary waves months later. Interestingly the variant shared in the photo above was correct in July 2011 but not in 2013, suggesting that perhaps memes have a more difficult time adapting as photos.

In conclusion, while much of the information that reaches us through our social networks is interesting and helpful, it doesn’t hurt to do a little bit of extra research before resharing that information.

Footnote:

¹ This sample is however heavily biased: given that the probability of receiving a comment mentioning Snopes is very small (between 0.1% and 0.3% of comments contain a link to Snopes), we are not only capturing a small fraction of rumors but we are also more likely to find larger cascades in our sample. To complicate matter further, the veracity of the rumor can also play a role in whether a share receives a Snopes comments (for example, false rumors elicit more Snopes links than true ones). While individual stories do vary in their likelihood of receiving a comment linking to Snopes, we can still do a rough estimate of the total number of rumors that are uploaded from the ones that do get detected that way — essentially, we infer the head of the distribution from its tail, assuming that the rate of receiving a Snopes link is accurate for cascades containing more than 10,000 shares.

This research was conducted using anonymized data.