Late last month, The New York Times announced that it was doing away with its public editor position, a sort of ombudsman for the paper, in part because the role had been filled by reader comments on the internet. The Times, publisher Arthur Sulzberger wrote in a memo, would be opening its comment sections on a greater number of stories and partnering with Google to manage them. It would be “a sea change in our ability to serve our readers, to hear from them, and to respond to them,” Sulzberger wrote. Today, the Times is unveiling the results of that partnership: an in-house process for automating comment moderation called Moderator.
Until recently, the New York Times manually approved 12,000 comments a day
Until recently, moderators for The New York Times would manually approve an average of 12,000 comments every day, with a team of 14 wading through reader responses as they came in. In a bid for a more targeted strategy, the Times partnered with Alphabet arm Jigsaw to organize comments based on their likelihood of being flagged for moderation. Times community editor Bassey Etim says the organization has already been using Moderator, and has been slowly ramping up the number of comments it approves through it. About 20 percent of comments are now automatically approved by the tool.
In February, Alphabet announced a related tool called Perspective, built to detect “toxic” comments. By mining millions of comments and drawing input from a human panel, the company said, it could begin to predict abusive text in a percentage format — a comment could be, for example, “40 percent similar to comments people said were ‘toxic.’” (Etim says the Times provided anonymized data to Alphabet for the project.) The system was by no means perfect: for a time, a demo seemed to identify all comments written in Arabic as partially toxic, even if the content of the text was uncontroversial. Sexist YouTube remarks plugged into the system by The Verge were marked as mostly non-toxic at the time. But less-nuanced comments, such as profanity-filled tirades, seemed easily caught in the net.
Like Perspective, Moderator charts comments based on their likelihood of being toxic or similar to comments previously flagged by moderators, creating a percentage score with the most likely to be flagged appearing closer to 100 percent. The Times will continue automatically approving some comments, and a ranked list will point moderators to the most controversial comments first. Etim says the organization won’t need to sacrifice too much accuracy for the sake of automation. As one example, he says, out of more than 8,300 opinion section comments submitted on a day last month, if the algorithm had been set to approve comments with a 10 percent or lower score, about 1,500 would be approved. Out of those 1,500, 22 would have been flagged by human moderators — not too far off from mistakes humans may have made anyway.
This new system, the Times says, will mean the organization can open up more stories to comments. Starting today, all of the organization’s “top stories” — those that appear with a summary on the Times homepage — during business hours will be opened for comments.
All of the organization’s “top stories” will be opened for comments
The Moderator announcement comes at a volatile time for comment sections around the internet. Several prominent news organizations have axed them entirely, arguing that the cost of moderation is not worth the potential rewards. Others have countered that comments are an essential reservoir of expertise, and that well-tended sections can build thoughtful audiences.
Meanwhile, in the internet at large, there are still questions about how massive social media audiences should be moderated, and for the most part, they rely heavily on human judgment. The Times is certainly a smaller, more special case — a paywalled news site — but the early success seems encouraging.
Still, there are hurdles in the Times’ quest to, as the guidelines put it, facilitate “civil” conversations that “avoid name-calling” — hurdles like teaching the algorithm when to flag newly coined insults about politicians. “There’s no historical record of ‘orange menace’ needing to be rejected,” Etim says.