We had caffe2 running a small model on the phone to try and select and propose photos for the user to share.
We were trying to offer an alternative sharing model that both made sharing easier, while offering the user the controls that made them feel comfortable with photo suggestions. (for those who never noticed, we launched Moments, which was an app that allowed automatic private sharing of your camera roll with a close selection of friends and family, but the experience wasn't great because it was centered around group events and sharing photos with the people who were there, not connecting with the ones who weren't)
Ultimately, it was scrapped, because we were paranoid that we hadn't come up with a user experience that made it clear that this was happening only on the phone (I think we even tried a notification model), or that we'd accidentally surface someone's boudoir photos, and we were too worried about the kind of knee-jerk reactions that you're seeing in this thread.
I'm guessing that someone at Meta either had a more successful go at the UX, or they feel that the opinions about AI have shifted enough that there will be less fear.
Upon reading the article, it looks like there are two options, one which is local-only, and similar to what we built, and a second one which tries to make better suggestions using online, and that is only enabled after asking the user.
I would suspect that the cloud processing version also runs a local model to attempt to filter out racy photos before sending them to the cloud, but I don't know for sure.
I think the article is a bit disingenuous in it's presentation, but it's possible that I'm biased because I know how a similar thing was built, but it definitely sounds like fear-mongering.