So the Tumblr adult content ban actually kicked in yesterday, December 17th, and I've spent the last twenty-four hours watching people's dashboards fill up with screenshots of the single dumbest content moderation failure I think I've ever seen in real time.
Quick recap for anyone who missed the announcement two weeks back: Tumblr said it was banning "adult content" starting the 17th, supposedly because Verizon (who owns them now, after the Yahoo/Oath mess) got spooked when the app got pulled from the App Store over child porn that was slipping past their existing filters. Fine, I get the panic. What I did not expect was for the actual enforcement mechanism to be this bad.
The classifier they built to flag "adult content" is, from what I can tell scrolling through my own feed this morning, mostly just looking at skin tones and shapes and guessing. People are posting screenshots of their own blogs getting flagged for things like a photo of a pair of boots, a close-up of a bowl of soup, and I saw one particularly good one where someone's post about knitting patterns got marked explicit. My personal favorite so far, and I wish I were making this up, was a screenshot of a public domain 19th century sculpture getting flagged, which is just deeply funny given the context of the whole thing being about "protecting" people.
I've had a Tumblr since around 2012, mostly for reposting stuff that didn't fit here on techpad, dumb GIF sets and half-finished thoughts I didn't want to bother formatting into a real post. I logged in this morning out of morbid curiosity and two of my own old posts, neither of which had anything remotely NSFW in them, one was literally a photo of my desk setup from like 2013, got the little "this post may contain sensitive content" banner slapped on it. There's an appeal process, apparently, but from what people are saying it's basically a black hole. You submit the appeal and then wait, and reviewers are apparently working through an enormous backlog by hand, which, when you think about the actual volume of content on that platform, is not a fast process.
The bigger story here, and the one I think actually matters more than the meme screenshots, is what this is doing to the artist and NSFW-adjacent communities that treated Tumblr as one of the last decent places to post that kind of work without getting immediately banned by an app store review team somewhere. A lot of those folks have spent the last two weeks scrambling to back up years of posts before the axe fell, and now that it has, plenty of them are just gone, migrated to Pillowfort or back to DeviantArt or wherever, sites that don't have anywhere near the audience Tumblr built up over the last decade plus. I don't think Tumblr's traffic ever recovers from this honestly, and I say that as someone who has watched this platform go through at least three separate "this is the year Tumblr dies" cycles since I started this blog back in 2011.
What gets me is that this was obviously rushed. You don't ship a machine learning content classifier this bad unless you had a deadline forced on you and didn't have time to actually tune the thing against a real dataset first. I'd bet good money there's an engineering team somewhere that flagged this exact problem internally and got overruled because legal wanted something, anything, live by the 17th no matter how it performed. That's usually how these things go. The deadline wins, the actual quality of the tool loses, and then a few million users get to deal with the fallout while the company issues a statement about how they're "continuing to refine the algorithm."
Anyway if your dashboard looks broken this week, that's why. Go dig through your own old posts, you'll probably find something dumb flagged too. Mine was a desk photo. Yours might be a picture of your cat, who knows anymore.