We’ve heard from a bunch of you that Safe Mode was filtering posts from the LGBTQ+ community even though they were completely innocuous and totally safe-for-work. Please know that was never our intention, and we appreciate you letting us know so quickly—and forcefully! We’re deeply sorry. Tumblr will always be a place where everyone is welcome and protected, so we want to explain what happened.
The major issue was some Tumblrs had marked themselves as Adult/NSFW (now Explicit) as a courtesy to their fellow users, and their perfectly safe posts were getting marked sensitive unintentionally. That should never have happened. We’re sorry.
We’re making some changes that should improve things:
Changes to self-marked blogs
What was happening: Because we consider Explicit blogs to be predominantly sensitive content, we were automatically marking all their posts as sensitive. That was too broad.
What we fixed: Now each post is classified individually. As they should be.
Changes to reblog chains
What was happening: If an Explicit Tumblr reblogged a safe post, we were marking that reblog as sensitive. This was even happening to text posts. Which is silly.
What we fixed: We changed the logic so that if the OP is safe, all its reblogs will also be safe.
Changes to photosets
What’s (still) happening: When you make a photo post, a computer algorithm classifies the image as safe or sensitive. It’s a machine so it’s not perfect. And the chances go up with photosets because there are multiple images. But out of an abundance of caution we keep posts marked sensitive until the OP requests a human review (by tapping the appeal button on their posts).
What we’re working on: We plan to have photosets analyzed as a whole group, rather than as individual images. That should reduce the number of mistakes the machine makes.
Safe Mode is supposed to make sure people aren’t surprised by things they may not want to see—specifically, nudity. It might take some time to get it perfect, but we’re committed to getting there with your help. Our algorithms will keep getting smarter as you give feedback on misclassified posts, and as you share your concerns and suggestions.