Couple of corrections because I got several things wrong: - credit where credit is due: the technology was developed by Microsoft; - it was developed as a gen

Couple of corrections because I got several things wrong: - credit where credit is due: the technology was developed by Microsoft; - it was developed as a generic image recognition software, hence the neutral name; - it is a hash, not a bloom filter; not sure why I remember it as a bloom; it doesn’t appear to be able to recognise new images; - the technology is widely used, but it was donated by Microsoft to International Centre for Missing & Exploited Children (ICMEC) which isn’t loosely connected but, yes, has close ties with all local law enforcement on those questions. It is available on Azure Marketplace, so widely available. The technology is still being expanded to work on videos and could be used to identify terrorism-related content. Because it’s meant to be resistant to alterations but not infer new content, it is not able to flag new content: that has to be manual. (Someone at the ICMEC has a horrifying job.) I would assume that it is generally implemented by flagging some images in uploaded albums and sending the whole album, or images from the same uploader, for manual verification. That would explain what I’ve heard about some images being potential. To come back to the original story: it seems more likely that Tumblr’s problems are related to new images.