If you’re a frequent user of Instagram, you’ve noticed some funny business in the comments recently. For example, comments the app deems most worthy or most relevant now defy chronological positioning. Behind the scenes, Instagram’s latest filters aren’t color treatments for photos—they’re content filters for mean and abusive comments. A new story in Wired details how Instagram built its new moderation tools, and their unlikely first test subject: Taylor Swift.
Last June, Instagram founder and CEO Kevin Systrom told Wired, he realized he wanted to make Instagram feel friendlier. The most technologically fascinating part of this effort uses Facebook’s proprietary DeepText machine-learning algorithm, which Instagram trained to filter comments much the way its team of human moderators already did. The system is designed to learn context over time, until—theoretically—it can discern an off-color inside joke from an outsider’s random insult. (As part of this effort, Instagram reportedly tags users with a hidden “karma score” based on past posts, a little like Uber’s “passenger rating.”)
If you tap to Settings > Options > Comments, you’ll find a toggle for “Hide Offensive Comments.” That’s Instagram’s program working for you. The other new moderation tool you can set up yourself: It’s called “Enable Keyword Filters,” and it allows users to plug in specific words they’d like to ban from comments on their posts. If you were Lena Dunham, for example, you’d probably want to ban the word “Lamby.”
The same feature allows users to ban particular emoji, and this is where Taylor Swift comes in. Last July, when Instagram was testing the new feature, Swift faced a wave of Kim Kardashian-induced backlash. Kanye West supporters accused Swift of duplicitousness, swarming her Instagram comments with snake emoji by the thousands. The first real-world test of Instagram’s customizable emoji ban wiped snakes off @taylorswift.