The most cited paper in this field is Soroush Vosoughi, Deb Roy, and Sinan Aral’s 2018 study in Science, “The spread of true and false news online.” It is cited, by now, for a great many things it does not actually say. Since the E in the FTSeffect rubric leans on it, this week is a Source Reading: what the paper measured, what it found, and where the citation chain has drifted from the text.
The study looked at roughly 126,000 cascades of contested stories on Twitter between 2006 and 2017, tweeted by about three million accounts. Truth values were not assigned by the authors. Each story was rated by six independent fact-checking organizations, and only stories on which those organizations agreed were kept. Within that set, false stories reached more people, faster, and were retweeted by more distinct users than true stories. The truth rarely diffused to more than one thousand people; the top one percent of false-news cascades routinely reached between one thousand and one hundred thousand. It took the truth about six times as long as falsehood to reach fifteen hundred people. The effect held across every category the authors examined, and was most pronounced in the political subset, which was the largest category in the data at roughly forty-five thousand cascades. Those are the headline numbers.
The paper did not measure the emotion of the sender. It measured the vocabulary of the reply.
The finding that gets pulled into the emotional-charge argument is more specific than the summary suggests. The authors coded the words in the replies to each cascade against the NRC Emotion Lexicon, a crowdsourced word list that tags English words for eight emotion categories. Replies to false stories used more words tagged as fear, disgust, and surprise. Replies to true stories used more words tagged as sadness, anticipation, joy, and trust. That is the entire emotional finding. It is a finding about the words that appeared in the responses, scored against a lexicon, not about what any individual reader felt and not about what any sender intended.
That distinction matters for the rubric. When the equation treats Emotional Charge as a variable, it is describing a property that the sender can dial up or down at the point of composition. The nearest study on the sender side is not Vosoughi. It is William Brady and colleagues in PNAS a year earlier, in 2017. Brady’s team scored the sender’s own text on moral-emotional vocabulary and found that each additional moral-emotional word was associated with roughly a twenty percent increase in retweets, with the effect largely contained inside ideological groups rather than crossing them. That is the sender-side result. Vosoughi is the reply-side result. Both belong in the chain; they are not the same claim.
A second drift is worth naming. Vosoughi is often paraphrased as showing that bots caused the differential spread of false news. It shows the opposite. The authors ran the analysis with and without accounts flagged as automated by two independent bot-detection systems. Bots accelerated the spread of true and false news at statistically indistinguishable rates. The differential was left after the bots were removed. The implication, which the paper states directly, is that the differential is a human phenomenon. It does not go away when the automation goes away.
A third drift is more subtle, and it is where careful readers of Duncan Watts, David Rothschild, and Markus Mobius have been useful correctives. False political news, in absolute terms, is a small share of the average American’s total information diet. The Vosoughi cascades are large as cascades, and the differential between true and false is real, but the base rate is small. This does not weaken the mechanism the paper identifies. It disciplines the story you are allowed to tell with it. Vosoughi shows that when a false story does move, it moves further and faster. It does not show that most of what moves is false.
Held to those boundaries, the paper still does exactly what the E variable needs it to do. It establishes, on a decade of data and against an outside truth standard, that novelty and affect are load-bearing in how contested content travels. The senders who understand that have priced it into their composition. The rest is measurement, and measurement, honestly reported, is what a rubric like the FTSeffect is trying to earn.
Next week: C is for Coordination, and the Kenya playbook.
— J.W.B.
Note 09 of 24