Transcription of SOCIAL SCIENCE The spread of true and false news online
1 SOCIAL SCIENCEThe spread of true and falsenews onlineSoroush Vosoughi,1 Deb Roy,1 Sinan Aral2*We investigated the differential diffusion of all of the verified true and false news storiesdistributed on Twitter from 2006 to 2017. The data comprise ~126,000 stories tweeted by~3 million people more than million times. We classified news as true or false usinginformation from six independent fact-checking organizations that exhibited 95 to 98%agreement on the classifications. Falsehood diffused significantly farther, faster, deeper, andmore broadly than the truth in all categories of information, and the effects were morepronounced for false political news than for false news about terrorism, natural disasters, SCIENCE , urban legends, or financial information. We found that false news was more novel thantrue news, which suggests that people were more likely to share novel information.
2 Whereasfalse stories inspired fear, disgust, and surprise in replies, true stories inspired anticipation,sadness, joy, and trust. Contrary to conventional wisdom, robots accelerated the spreadof true and false news at the same rate, implying that false news spreads more than the truthbecause humans, not robots, are more likely to spread theories of decision-making(1 3), cooperation (4), communication (5),and markets (6) all view some concep-tualization of truth or accuracy as centralto the functioning of nearly every humanendeavor. Yet, both true and false informationspreads rapidly through online media. Definingwhat is true and false has become a commonpolitical strategy, replacing debates based ona mutually agreed on set of facts. Our economiesare not immune to the spread of falsity rumors have affected stock prices and themotivation for large-scale investments, for ex-ample, wiping out $130 billion in stock valueafter a false tweet claimed that Barack Obamawas injured in an explosion (7).
3 Indeed, our re-sponses to everything from natural disasters(8,9) to terrorist attacks (10)havebeendisruptedby the spread of false news SOCIAL technologies, which facilitate rapidinformation sharing and large-scale informationcascades, can enable the spread of misinformation( , information that is inaccurate or misleading).But although more and more of our access toinformation and news is guided by these newtechnologies (11), we know little about their con-tribution to the spread of falsity online . Thoughconsiderable attention has been paid to anecdotalanalyses of the spread of false news by the media(12), there are few large-scale empirical investiga-tions of the diffusion of misinformation or its socialorigins. Studies of the spread of misinformationare currently limited to analyses of small, ad hocsamples that ignore two of the most importantscientific questions: How do truth and falsitydiffuse differently, and what factors of humanjudgment explain these differences?
4 Current work analyzes the spread of singlerumors, like the discovery of the Higgs boson(13) or the Haitian earthquake of 2010 (14), andmultiple rumors from a single disaster event, likethe Boston Marathon bombing of 2013 (10), or itdevelops theoretical models of rumor diffusion(15), methods for rumor detection (16), credibilityevaluation (17,18), or interventions to curtail thespread of rumors (19).Butalmostnostudiescom-prehensively evaluate differences in the spreadof truth and falsity across topics or examinewhy false news may spread differently than thetruth. For example, although Del Vicarioet al.(20) and Bessiet al.(21) studied the spread ofscientific and conspiracy-theory stories, theydid not evaluate their veracity. Scientific andconspiracy-theory stories can both be either trueor false , and they differ on stylistic dimensionsthat are important to their spread but orthogonalto their veracity.
5 To understand the spread offalse news, it is necessary to examine diffusionafter differentiating true and false scientific storiesandtrueandfalseconspiracy-theorys toriesandcontrolling for the topical and stylistic differencesbetween the categories themselves. The only studyto date that segments rumors by veracity is that ofFriggeriet al.(19), who analyzed ~4000 rumorsspreading on Facebook and focused more on howfact checking affects rumor propagation than onhow falsity diffuses differently than the truth (22).In our current political climate and in theacademic literature, afluidterminology has arisenaround fake news, foreign interventions politics through SOCIAL media, and our under-standing of what constitutes news, fake news, false news, rumors, rumor cascades, and otherrelated terms. Although, at one time, it may havebeen appropriate to think of fake news as refer-ring to the veracity of a news story, we nowbelieve that this phrase has been irredeemablypolarized in our current political and media cli-mate.
6 As politicians have implemented a politicalstrategy of labeling news sources that do notsupport their positions as unreliable or fake news,whereas sources that support their positions arelabeled reliable or not fake, the term has lost allconnection to the actual veracity of the informa-tion presented, rendering it meaningless for usein academic classification. We have therefore ex-plicitly avoided the term fake news throughoutthis paper and instead use the more objectivelyverifiable terms true or false news. Althoughthe terms fake news and misinformation alsoimply a willful distortion of the truth, we do notmakeanyclaimsabouttheintentofthepurve yorsof the information in our analyses. We insteadfocus our attention on veracity and stories thathave been verified as true or also purposefully adopt a broad definitionof the term news. Rather than defining whatconstitutes news on the basis of the institutionalsource of the assertions in a story, we refer to anyasserted claim made on Twitter as news (we de-fend this decision in the supplementary materialssection on reliable sources, section ).
7 Wedefine news as any story or claim with an asser-tion in it and a rumor as the SOCIAL phenomenaof a news story or claim spreading or diffusingthrough the Twitter network. That is, rumors areinherently SOCIAL and involve the sharing of claimsbetween people. News, on the other hand, is anassertion with claims, whether it is shared or rumor cascade begins on Twitter when auser makes an assertion about a topic in a tweet,which could include written text, photos, or linksto articles online . Others then propagate therumor by retweeting it. A rumor sdiffusionpro-cess can be characterized as having one or morecascades, which we define as instances of a rumor-spreading pattern that exhibit an unbroken re-tweet chain with a common, singular origin. Forexample, an individual could start a rumor cas-cade by tweeting a story or claim with an assertionin it, and another individual could independentlystart a second cascade of the same rumor (per-taining to the same story or claim) that is com-pletely independent of thefirst cascade, exceptthat it pertains to the same story or claim.
8 If theyremain independent, they represent two cascadesof the same rumor. Cascades can be as small as sizeone (meaning no one retweeted the original tweet).Thenumberofcascadesthatmakeuparum orisequal to the number of times the story or claim wasindependently tweeted by a user (not retweeted).So, if a rumor A is tweeted by 10 people separate-ly, but not retweeted, it would have 10 cascades,each of size one. Conversely, if a second rumor B is independently tweeted by two people andeach of those two tweets is retweeted 100 times,the rumor would consist of two cascades, eachof size we investigate the differential diffusionof true, false , and mixed (partially true, partiallyfalse) news stories using a comprehensive dataset of all of the fact-checked rumor cascades thatspread on Twitter from its inception in 2006 to2017. The data include ~126,000 rumor cascadesspreadby~ We sampled all rumor cascades investigatedby six independent fact-checking organizationsRESEARCHV osoughiet al.
9 ,Science359, 1146 1151 (2018)9 March 20181of61 Massachusetts Institute of Technology (MIT), the Media Lab,E14-526, 75 Amherst Street, Cambridge, MA 02142, ,E62-364,100 MainStreet,Cambridge,MA02142,USA.*Corres ponding author. Email: August 30, 2020 from ( , , , , , and ) by parsing the title, body, and verdict(true, false , or mixed) of each rumor investigationreported on their websites and automaticallycollecting the cascades corresponding to thoserumors on Twitter. The result was a sample ofrumor cascades whose veracity had been agreedon by these organizations between 95 and 98% ofthe time. We cataloged the diffusion of the rumorcascades by collecting all English-language repliesto tweets that contained a link to any of theaforementioned websites from 2006 to 2017 andused optical character recognition to extract textfrom images where needed.
10 For each reply tweet,we extracted the original tweet being replied toand all the retweets of the original tweet. Eachretweet cascade represents a rumor propagatingon Twitter that has been verified as true or falsebythefact-checkingorganizations(see thesup-plementary materials for more details on cascadeconstruction). We then quantified the cascades depth (the number of retweet hops from theorigin tweet over time, where a hop is a retweetby a new unique user), size (the number of usersinvolved in the cascade over time), maximumbreadth (the maximum number of users involvedin the cascade at any depth), and structural vi-rality (23) (a measure that interpolates betweencontent spread through a single, large broadcastand that which spreads through multiple gen-erations, with any one individual directly respon-sible for only a fraction of the total spread ) (seethe supplementary materials for more detail onthe measurement ofrumor diffusion).
