新型闢謠網站：實時追蹤網絡流言散佈軌跡

Did you hear the thing about the Florida woman who implanted a third breast in order to be "unattractive to men"? The one who is filming "her daily life in Tampa to show the struggles she faces because of her surgery"?

She didn't, and she is not. The whole thing—for better or for worse—was a hoax.

The Internet moves quickly. Rumors emerge, intentionally and not; they spread, intentionally and not. There's a reason, of course, that "wildfire" is such a common metaphor when it comes to describing this stuff: Rumors, once sparked, don't just spread extremely quickly; they are also extremely difficult to contain. And on top of everything else, it is extremely hard to predict which direction they'll take as they spread.

Enter . The site, launched today after two months of testing and data-gathering, is hoping to change that by tracking rumors that arise in (pretty much) real time. As Craig Silverman, the rumor researcher who created the site, told me: "It's aiming to be a real-time monitoring of claims that are emerging in the press."

works through a combination of human and algorithmic processing. Silverman and a research assistant find rumors that are being reported in the mainstream media—most often, stories that bubble up through social media and get spotted by one outlet … and then, from there, picked up by other outlets. Then they search Google News, which aggregates various news outlets' take on the story. They gather those stories and enter them into their database, classifying them according to the outlet and to how each of the outlets is reporting them. Some will report rumors not as rumors, but rather as simply true or simply false; the majority, however, simply report the fact that they have heard them. Often they'll hedge that repetition with caveats like "Sources:" or "Rumor:" or "Unconfirmed:" in headlines or texts. Just as often, however, they'll be more subtle in their warnings. In the case of the Tri-Breasted Lady, many places simply repeated the rumor as fact, their main additional caveat being a well-placed "WTF."

Then their algorithm takes over. is essentially a web app, built on an API from the back-end database. (You can think of it as something of a data-driven version of Snopes—with a more expansive premise. "Snopes, they're amazing," Silverman says, "but they only do Snopes work; they don't aggregate what other people are doing. This site is able to identify claims and then actually see, okay, who's got the best information about it?") Every hour, 's script crawls the stories to see whether their text has changed. The system also checks share counts to monitor how stories are moving through the social media ecosystem.

All of which, Silverman says, helps to answer questions that haven't been so systematically analyzed before. Among them: "What's the life cycle of a rumor in the press now? And how are news organizations dealing with things that are unconfirmed? And are they updating the stories, and are they sticking with it over time?"

Take the reports—false reports, it turns out—that Durex is making a pumpkin-spice condom: Through , you can see who repeated the rumor, who checked it, and who debunked it. Or take, more seriously, the stories that emerged last week claiming that a meteorite had landed in Nicaragua. (That rumor is still listed as unverified on , because no one has been able to prove that such a meteorite actually fell.)

A challenge news organizations face when it comes to rumor-reporting in particular is the fact that rumors tend to be much more shareable—and much more clickable—than corrections. Take your friend and mine, Ms. Tampa Triple-Breast (self-given pseudonym: Jasmine Tridevil). One of the early stories about her, in the New York Post, got 40,000 shares. The Snopes article (mostly) debunking the initial story had 12,500 shares—a decent amount for a story that is, technically, a non-story.

For the most part, though, the articles debunking the rumor get extremely little attention. (For a more crystalline example of all that, you can look to Buzzfeed's coverage of the story. Its initial story got more than 30,000 shares; its debunking of that story got just over 1,000.) Which means that news organizations often have very little incentive—direct, commercial incentive, at least—to put their time and energy into them. As a result, as Silverman puts it: The Total Recall rumor is "a story that, I would argue, the average person probably doesn't know is not true."

The larger problem with all that is that rumors, once they're put out there into the maw of the media, are notoriously hard to correct. There's the fact that "sorry, just kidding about that three-boobed lady thing" is nowhere near as sharable as a "whoa, three-boobed lady!" thing in the first place. But there's also the fact that there is very little uniformity among media outlets about how updates, corrections, retractions, and the like should be presented to readers. Most outlets will simply update a story that contains a debunked claim; a few will write new stories altogether, linking to the previous one in the process. That can leave readers, however, in a kind of epistemological limbo: You're never quite sure what's been verified and what has not. Trust is a precious resource in journalism; many outlets haven't fully figured out how to preserve it.

"So much of this stuff is public before news organizations get to it," Silverman points out. "So that's a very different dynamic from what used to happen. So if something is by default public, how do you decide when you're doing to point at it in a way that's responsible? And then how do you deal with it as it sort of takes its life path to being true or false?" Bringing some data to bear on those questions, he's hoping, will help news outlets start to answer them.據《大西洋月刊》雜誌網站報道，你聽說過那個有三個乳房的佛羅里達州女人嗎？那個爲了讓自己“在男人面前沒有吸引力”而去移植第三個乳房，還拍攝了自己“在坦帕市（Tampa）的日常生活，及由於手術而遭受的各種掙扎”錄像的女人——你聽說過沒？

她沒有做過那種手術，她也沒有第三個乳房。整件事不論怎樣——是一場騙局。

網上消息傳播快。不論有無意，謠言空穴起；不論有無心，謠言傳千里。這當然有原因，“野火”是形容這種事的常用隱喻：謠言，一旦被點燃，不僅流傳極快，而且極難控制。不過最可怕的是，我們極難預知謠言會往哪個方向發展。

登錄網站。經過兩個月的測試和收集數據，這個網站於9月28日正式啓用，希望通過追蹤流言產生的（接近）真實時間，改變流言四起的困境。網站創立者，謠言研究員克雷格•西爾弗曼（Craig Silverman）告訴我：“網站致力於實時監測媒體中產生的各種言論。”

在比較廣大的前提下，你可以把它當作數據驅動版的Snopes（美國一家專門覈查並揭穿謠言和傳聞的網站）。

網站通過人工處理和算法處理相結合來運作。西爾弗曼和一位研究助理髮現，那些被主流媒體報道的謠言大多最早只是在社交媒體中冒了個泡，而後被某個渠道發現……然後以此爲中轉，被其他渠道獲取了消息。他們隨後搜索了谷歌新聞，這裏聚集各類新聞渠道對消息的報道。他們將這些報道收集編入數據庫，並以其渠道來源和報道方式進行分類。有的報道不把謠言當謠言，而是進行了簡單的真或假區分；不過大部分報道僅僅告訴讀者有這麼個傳言。通常標題或正文中會用一些像“來源”，“傳言”，或“未確認”之類的標註來避免重複。然而，往往報道中的警告會更加微妙。在“三乳女子”事件中，許多媒體僅僅將謠言當作事實重複，它們的主要附加警告便是一個方便順手的“WTF.”

他們的計算機算法接手了這種狀況。實質是一個網頁應用，建立於一個後端數據庫的應用程序界面（API）。（在比較廣大的前提下，你可以把它當作數據驅動版的Snopes。“Snopes非常棒。”西爾弗曼表示，“但是它也有侷限，它無法收集其他人正在做什麼的信息。這個網站能夠識別各種言論，然後切實發現是誰掌握了消息的最可靠信息。”）每隔一小時，的腳本程序就會監測消息的文字是否發生改變。這個系統還檢查消息被分享的次數，以此監控信息在社交媒體生態系統中是如何移動的。

所有這些，西爾弗曼表示，有助於回答不少此前從未被系統分析過的問題。其中包括：“謠言在新聞界的生命週期是什麼樣的？新聞機構如何處理未經確定的消息？報道是否有及時更新，或是一直不變？”

拿那些報道——假報道爲例，竟有“杜蕾斯（Durex）推出南瓜香型的安全套”的消息：通過，你可以分別知道是誰散佈、遏制以及揭穿了謠言。或者以較嚴肅的報道爲例，上週有數篇報道稱有一顆隕石墜落於尼加瓜拉。（這則謠言依然在列於的未經證實列表中，因爲無人可以證明確有這麼一顆隕石墜落。）

新聞機構尤其在報道謠言時會面臨一個挑戰：謠言比闢謠更容易被分享——也更容易吸引點擊。以大家都熟悉的“坦帕市三乳女士”爲例（她給自己造了假名字：茉莉•三乳惡魔（Jasmine Tridevil））。早期一篇《紐約郵報》（New York Post）關於她的報道被分享了4萬次。而Snopes發表的（主要）拆穿初期報道的文章僅1.25萬次分享——這對於一篇從技術層面上不算報道的報道來說，也相當可觀。

“三乳妖婦，純屬惡搞”報道遠不及“亮瞎！驚現三乳妖婦”報道來得吸睛。

然而大多情況下，闢謠的文章總是乏人問津。（舉個最清楚透明的例子，你可以查看在新聞聚合網站Buzzfeed對那個新聞的報道。首篇文章被分享逾3萬次；而闢謠文章僅1萬餘次。）這意味着新聞機構缺乏足夠的激勵——直接的商業激勵，懶得把時間和精力放在闢謠上。結果如西爾弗曼所說的：電影《全面回憶》（Total Recall）的謠言就是“一個我敢說大多數人都信以爲真的故事”。

更大的問題在於，謠言一經媒體的輪番反芻，便人盡皆知，難以澄清。最近的例子就是“抱歉，三乳妖婦，純屬惡搞”的分享量遠不及“亮瞎！驚現三乳妖婦”。不僅如此，各媒體渠道在如何向讀者更新、澄清、撤回文章等問題上缺乏默契。大多數媒體僅簡單更新一篇闢謠聲明的報道；鮮有媒體會集中重新報道，並提供先前文章的鏈接。如此便給讀者留下了認識上的不安定狀態：你總是無法確定什麼是已經證實的事實，什麼是無憑無據的謠言。信任是新聞業的寶貴資源，但許多媒體還不知道要如何保護它。

“這些東西不少在經手媒體之前就已經衆人皆知，”西爾弗曼指出，“所以這是一種與以往極爲不同的動力。因此，如果面對一樁已經公衆化的事情，你決定以怎樣負責任的方式處理？如果它正沿着其或真或假的軌道發展，你又該如何處置？”就此，西爾弗曼收集了一些數據，他希望可以藉此幫助新聞機構回答這些問題。