Less than you think: Prevalence and predictors of fake news dissemination on Facebook

Less than you think: Prevalence and predictors of fake news dissemination on Facebook Andrew M. Guess Jonathan Nagler Joshua A. Tucker

Tag layer of the DHSS Meta-Tagging Project. The article text is reproduced from the staged copy named in sourceDesc and is not altered; the tags are held separately in standOff so that overlapping and nested spans are preserved.

Less than you think: Prevalence and predictors of fake news dissemination on Facebook 2019 Science Advances 10.1126/sciadv.aau4586 guess2019-less-than-you-think-prevalence-and guess2019-less-than-you-think-prevalence-and.txt, the project's staged copy cd4bb8aada0f7b18bd6bc9b9cb8bfd336844740c cc-by-nc political-science

the paper's own words

the tag layer; a sentence written by the tagger about the paper, always anchored to a quotation but not itself a quotation

0 spans located by searching the stored text for the quote's exact words, allowing any whitespace between them who actually posted links from fake news domains to their Facebook friends during the 2016 US campaign, measured from profile data rather than from what people say they did Quotation withheld because this record is not cleared for redistribution. how common fake news sharing was, and which individual characteristics predict it once ideology, partisanship, education and overall posting volume are held constant Quotation withheld because this record is not cleared for redistribution. a survey of 3500 linked to 1191 respondents' real Facebook sharing: over 90% shared no fake news at all, conservatives shared more, and the oldest group shared nearly seven times as much as the youngest with everything else held constant Quotation withheld because this record is not cleared for redistribution. existence Quotation withheld because this record is not cleared for redistribution. the dissemination of deliberately false political content by ordinary users - rare, concentrated, and predicted by age more robustly than by anything else measured Quotation withheld because this record is not cleared for redistribution. observational-study observational-study Quotation withheld because this record is not cleared for redistribution. BuzzSumo ranks domains by Facebook engagement; a journalist checks each for the hallmark features of a fake news site - no contact page, syndicated content, recently registered; a supervised classifier trained on 7 million pages removes anything it calls hard news. Thirty domains become twenty-one, and every share in the dependent variable is a link to one of those twenty-one a category built by three procedures in sequence Quotation withheld because this record is not cleared for redistribution. linking profile data to a survey removes the self-report bias in what people say they shared, and it cannot recover what they thought they were sharing. The design buys accuracy about the act and gives up access to the state of mind - which is the quantity the survey experiments measure and cannot observe the act of behaviour without belief Quotation withheld because this record is not cleared for redistribution. 3500 respondents, 1331 consenting to share profile data, 1191 successfully linked - and the linked respondents were more knowledgeable and more politically engaged than those who declined. The behavioural half of the design is available only from people willing to hand it over, and willingness correlates with the subject matter a consent gradient inside the sample Quotation withheld because this record is not cleared for redistribution. sharing is a choice made over what arrived. Without the News Feed, a group that shares more fake news may be a group that saw more of it, and no amount of control for the sharer's characteristics separates the two. The authors state this rather than resolving it the unobserved half of every social media finding Quotation withheld because this record is not cleared for redistribution. there is no measure of whether respondents knew that what they shared was false Quotation withheld because this record is not cleared for redistribution. the data are observational and cannot provide causal evidence, and an omitted variable cannot be definitively ruled out Quotation withheld because this record is not cleared for redistribution. half the respondents with Facebook accounts declined to share their profile data, and the linked ones were more knowledgeable and politically engaged Quotation withheld because this record is not cleared for redistribution. the composition of respondents' News Feeds is unobserved, so exposure cannot be separated from willingness to share Quotation withheld because this record is not cleared for redistribution. posts deleted before consent are invisible, and the authors argue rather than measure that this should reduce noise rather than bias Quotation withheld because this record is not cleared for redistribution. the conservative finding is sensitive to coding and rests on a small number of respondents Quotation withheld because this record is not cleared for redistribution. the ideological pattern may reflect the pro-Trump slant of the available content and the ordinary tendency to share what one agrees with, rather than a greater propensity to share falsehood Quotation withheld because this record is not cleared for redistribution. there is no commonly accepted ground-truth standard for what counts as fake news, so the dependent variable rests on lists built by journalists and academics Quotation withheld because this record is not cleared for redistribution. digital trace data linked to survey response, carried from computational social science into a question political science had been answering with self-report Quotation withheld because this record is not cleared for redistribution. a supervised text classifier from information retrieval, used not to make the finding but to draw the boundary of the outcome variable Quotation withheld because this record is not cleared for redistribution. the total number of links shared of any kind reported beside the fake news count, and included as a covariate in every model eliminates: a low fake-news count meaning people do not share links - most respondents shared hundreds, so the rarity is about the content · outcome: held Quotation withheld because this record is not cleared for redistribution. the relationship between overall sharing volume and fake news sharing examined directly, rather than only controlled for eliminates: the explanation that some respondents will share anything - the heaviest sharers shared LESS fake news, which is the opposite of that prediction · outcome: held Quotation withheld because this record is not cleared for redistribution. the same model applied to hard news sharing as a placebo outcome eliminates: age predicting online sharing in general - if it did, the predictors of hard news would look the same, and they do not include age at all · outcome: held Quotation withheld because this record is not cleared for redistribution. five different lists of fake news domains, including one built from 948 individual fact-checks and one crowdsourced, with the analysis rerun on each eliminates: one journalist's list carrying the finding - the category has no ground truth, so the robustness has to be to the choice of vocabulary itself · outcome: held Quotation withheld because this record is not cleared for redistribution. an exact-URL variant of the analysis using only the 948 fact-checked stories, rather than counting every article from a listed domain eliminates: the assumption that everything published by a fake news domain is fake - the stricter definition does not need it · outcome: held Quotation withheld because this record is not cleared for redistribution. hard news domains removed from the list by a classifier trained on 7 million shared pages, so that hyperpartisan sites are excluded from the outcome eliminates: the contested hyperpartisan category being counted as fake news, which would make an ideological finding partly definitional · outcome: held Quotation withheld because this record is not cleared for redistribution. negative binomial and OLS models run alongside the quasi-Poisson, with the quasi-Poisson chosen by a dispersion test eliminates: the count model's distributional assumptions producing the result · outcome: held Quotation withheld because this record is not cleared for redistribution. ideology and party identification, correlated at r = 0.31, entered separately and together, with all three specifications reported eliminates: a collinearity decision being made silently in the authors' favour - they name the correlation and publish every version · outcome: held Quotation withheld because this record is not cleared for redistribution. the models re-fitted as negative binomial and OLS alongside the Poisson and quasi-Poisson specifications, with dispersion tests deciding between the latter two eliminates: that the sharing result is an artefact of one count-model specification · outcome (stated by the tagger, no quotation): held Quotation withheld because this record is not cleared for redistribution. respondents asked to check which categories of profile information they were willing to share, with no access to News Feeds or to friends' data eliminates: collecting more than was consented to - the field-by-field opt-in bounds the data at the point of collection · outcome (stated by the tagger, no quotation): held Quotation withheld because this record is not cleared for redistribution. fake news fake news Quotation withheld because this record is not cleared for redistribution. hard news hard news Quotation withheld because this record is not cleared for redistribution. links links Quotation withheld because this record is not cleared for redistribution. age age Quotation withheld because this record is not cleared for redistribution. ideology ideology Quotation withheld because this record is not cleared for redistribution. News Feeds News Feeds Quotation withheld because this record is not cleared for redistribution. quasi-Poisson quasi-Poisson Quotation withheld because this record is not cleared for redistribution. digital media literacy digital media literacy Quotation withheld because this record is not cleared for redistribution. fake news sharing - AS MEASURED HERE: a post to one's own timeline whose link resolves to one of 21 domains, from profile data, with no measure of whether the sharer believed it Quotation withheld because this record is not cleared for redistribution. fake news - AS DEFINED: knowingly false or misleading content created largely to generate ad revenue, with the authors stating there is no commonly accepted ground-truth standard for the category Quotation withheld because this record is not cleared for redistribution. the domain list - AS BUILT BY BUZZFEED: 30 domains ranked by Facebook engagement and checked by a journalist for the hallmark features of a fake news site Quotation withheld because this record is not cleared for redistribution. the domain list - AFTER THE HARD-NEWS FILTER: 21 domains, with hyperpartisan sites such as Breitbart removed by a supervised classifier so that the contested category is excluded Quotation withheld because this record is not cleared for redistribution. the conservative effect - AS AN OBSERVATION: conservatives shared the most, and the content was overwhelmingly pro-Trump Quotation withheld because this record is not cleared for redistribution. the conservative effect - AS AN INFERENCE: possibly not a greater tendency to share fake news at all, but the ordinary tendency to share what one agrees with, conditional on exposure the study cannot see Quotation withheld because this record is not cleared for redistribution. critiques → post-election narratives about the role of social media disinformation, by measuring how rare the behaviour actually was - over 90% of linked respondents shared none at all Quotation withheld because this record is not cleared for redistribution. critiques → self-reported measures of online activity, which the linked design is built to avoid Quotation withheld because this record is not cleared for redistribution. critiques → the discipline's habit of treating demographic traits as controls to be conditioned away rather than as explanatory factors in their own right Quotation withheld because this record is not cleared for redistribution. qualifies → its own ideological finding - sensitive to coding, based on few respondents, and confounded with exposure the design cannot observe Quotation withheld because this record is not cleared for redistribution. fake news —distinct-from→ hard news fake news hard news differs in: whether the site is intentionally or systematically factually inaccurate, or merely partisan · the relation itself is separately evidenced · strength of the relation: 0.95 · differs in (stated by the tagger, no quotation): whether the site is intentionally or systematically factually inaccurate, or merely partisan · the relation stated: domains a supervised classifier calls hard news are removed, so that the contested hyperpartisan category is not counted as fake news Quotation withheld because this record is not cleared for redistribution. age —distinct-from→ ideology age ideology differs in: whether the predictor survives conditioning on the other · the relation itself is separately evidenced · strength of the relation: 0.95 · differs in (stated by the tagger, no quotation): whether the predictor survives conditioning on the other · the relation stated: holding ideology, party or both constant, each age group still shared more than the next-youngest Audit 2026-09-28: DOUBTFUL - read against its span and kept, but age effects hold with ideology held constant; distinct-from is loose. Quotation withheld because this record is not cleared for redistribution. over 90% of linked respondents shared no articles from fake news domains during the 2016 campaign over 90% of linked respondents shared no articles from fake news domains during the 2016 campaign Quotation withheld because this record is not cleared for redistribution. the rarity is specific to the content - most respondents shared between 100 and 1000 links of other kinds the rarity is specific to the content - most respondents shared between 100 and 1000 links of other kinds Quotation withheld because this record is not cleared for redistribution. those over 65 shared nearly seven times as many fake news articles as the youngest group, holding ideology, education and posting volume constant those over 65 shared nearly seven times as many fake news articles as the youngest group, holding ideology, education and posting volume constant Quotation withheld because this record is not cleared for redistribution. the heaviest sharers overall shared LESS fake news, so the behaviour is not explained by a general willingness to share anything the heaviest sharers overall shared LESS fake news, so the behaviour is not explained by a general willingness to share anything Quotation withheld because this record is not cleared for redistribution. the same model applied to hard news sharing gives varied predictors and does not include age the same model applied to hard news sharing gives varied predictors and does not include age Quotation withheld because this record is not cleared for redistribution. The fake-news list contains 21 mostly pro-Trump domains. The fake-news list contains 21 mostly pro-Trump domains. Quotation withheld because this record is not cleared for redistribution. the results are essentially invariant across five different lists of what counts as fake news the results are essentially invariant across five different lists of what counts as fake news Quotation withheld because this record is not cleared for redistribution. there is no measure of whether respondents knew what they shared was false there is no measure of whether respondents knew what they shared was false Quotation withheld because this record is not cleared for redistribution. knowingly false or misleading content created largely for ad revenue - with the authors stating that the term covers a variety of phenomena and that no commonly accepted ground-truth standard exists fake news Quotation withheld because this record is not cleared for redistribution. of linked respondents shared no articles from fake news domains at all - the number the title refers to 91.5% Quotation withheld because this record is not cleared for redistribution. web domains constituting the entire dependent variable, after a journalist's list of 30 was filtered by a hard-news classifier 21 value source note: THE DESCRIPTION DRAWS ON THE PAPER BEYOND THIS QUOTE: the starting list of 30 is described in the appendix ('Our BuzzFeed-based list began with 30 domains'). Quotation withheld because this record is not cleared for redistribution. of the 3500 survey respondents agreed to share Facebook profile data; 1191 were successfully linked 38% Quotation withheld because this record is not cleared for redistribution. the correlation between ideological self-placement and party identification, disclosed and handled by reporting all three specifications 0.31 Quotation withheld because this record is not cleared for redistribution. quasi-Poisson quasi-Poisson Quotation withheld because this record is not cleared for redistribution. CI CI Quotation withheld because this record is not cleared for redistribution. age age Quotation withheld because this record is not cleared for redistribution. fake news fake news Quotation withheld because this record is not cleared for redistribution.

The full text of this article is not redistributable, so it is not included here. The evidence sentences are withheld with it; the tag layer and its grounding status remain. The article itself is at https://doi.org/10.1126/sciadv.aau4586.