In August 2018, President Donald Trump claimed that social media was “completely discriminating towards Republican/Conservative voices.” Not a lot was new about this: for years, conservatives have accused tech corporations of political bias. Simply final July, Senator Ted Cruz (R-Texas) requested the FTC to research the content material moderation insurance policies of tech corporations like Google. A day after Google’s vice chairman insisted that YouTube was apolitical, Cruz claimed that political bias on YouTube was “large.”
However the information does not again Cruz up—and it has been obtainable for some time. Whereas the precise insurance policies and procedures for moderating content material are sometimes opaque, it’s potential to take a look at the outcomes of moderation and decide if there’s indication of bias there. And, final yr, pc scientists determined to do precisely that.
Motivated by the long-running argument in Washington DC, pc scientists at Northeastern College determined to research political bias in YouTube’s remark moderation. The staff analyzed 84,068 feedback on 258 YouTube movies. At first look, the staff discovered that feedback on right-leaning movies appeared extra closely moderated than these on left-leaning ones. However when the researchers additionally accounted for components such because the prevalence of hate speech and misinformation, they discovered no variations between remark moderation on right- and left-leaning movies.
“There is no such thing as a political censorship,” mentioned Christo Wilson, one of many co-authors and affiliate professor at Northeastern College. “In reality, YouTube seems to simply be implementing their insurance policies towards hate speech, which is what they are saying they’re doing.” Wilson’s collaborators on the paper had been graduate college students Shan Jiang and Ronald Robertson.
To verify for political bias in the best way feedback had been moderated, the staff needed to know whether or not a video was right- or left-leaning, whether or not it contained misinformation or hate speech, and which of its feedback had been moderated over time.
From fact-checking web sites Snopes and PolitiFact, the scientists had been in a position to get a set of YouTube movies that had been labelled true or false. Then, by scanning the feedback on these movies twice, six months aside, they may inform which of them had been taken down. Additionally they used pure language processing to determine hate speech within the feedback.
To assign their YouTube movies left or proper scores, the staff made use of an unrelated set of voter information. They checked the voters’ Twitter profiles to see which movies had been shared by Democrats and Republicans and assigned partisanship scores accordingly.
The uncooked numbers “would appear to recommend that there’s this form of imbalance by way of how the moderation is going on,” Wilson mentioned. “However then once you dig somewhat deeper, when you management for different components just like the presence of hate speech and misinformation, rapidly, that impact goes away, and there is an equal quantity of moderation occurring within the left and the fitting.”
Kristina Lerman, a pc scientist on the College of Southern California, acknowledged that research of bias had been tough as a result of the identical outcomes may very well be brought on by various factors, identified in statistics as confounding variables. Proper-leaning movies could merely have attracted stricter remark moderation as a result of they bought extra dislikes or contained faulty data or as a result of the feedback contained hate speech. Lerman mentioned that Wilson’s staff had factored various prospects into their evaluation utilizing a statistical technique generally known as propensity rating matching and that their evaluation appeared “sound.”
Kevin Munger, a political scientist at Penn State College, mentioned that, though such a examine was necessary, it solely represented a “snapshot.” Munger mentioned that it will be “rather more helpful” if the evaluation may very well be repeated over an extended time period.
Within the paper, the authors acknowledged that their findings could not be generalized over time as a result of “platform moderation insurance policies are notoriously fickle.” Wilson added that their findings could not be generalized to different platforms. “The massive caveat right here is we’re simply YouTube,” he mentioned. “It might be nice if there was extra work on Fb, and Instagram, and Snapchat, and no matter different platforms the children are utilizing today.”
Wilson additionally mentioned that social media platforms had been caught in a “deadly embrace” and that each determination they made to censor or enable content material was certain to attract criticism from the opposite aspect of the political spectrum.
“We’re so closely polarized now—possibly nobody will ever be pleased,” he mentioned with fun.