They say they the images are merely matched to pre-determined images found on the web. You're talking about a different scenario where AI detects inappropriate contents in an image.

So if I send a photo of our kids playing naked in a baby pool to my wife through signal, some slimy-ass eurocrat in some IT center will be able to 'manually verify' the photo of my naked kids?? Are you mentally sound??

I hate pedos just as much as every other sane parent, perhaps even more so (I'd love to wear "Why, Garry, why?!" t-shirt all day every day). But to hell with this stupid idea that some slimy scumbags will be able to browse my own photos of my own kids. Hell, even any random photo I take, it's my business and nobody elses! Go catch pedos the proper way instead, work a little, we don't need Gestapo or Stasi to hover over everything we do or photograph.

technically it doesn't break e2ee

** for some unorthodox definition of e2ee

If the "endpoints" are defined as being somewhere outside the end users' control, because for example the client software they have is designed to betray their secrets, then the system is no longer end-to-end encrypted in the way that both cryptographers and normal people would usually understand the concept.

I think the parent is distinguishing between messages & the attachments as they are stored differently & often in different places in many systems. But I agree with you in assuming that the goal would ultimately be to then start scanning messages too.

Imagine governments used something like SHA1 that has conflicts & now you have collision potential--you could even fabricate attachments that could cause a collision to get someone throw in jail since all you have to rely on is the file hashes. If you can’t scan the actually content & you are just using hashes, then you also don’t prevent new content that those in power deem ‘bad’ from being flagged either which doesn’t really stop the proliferation of the ‘bad thing’ just specific known ‘bad things’. If I were implementing clients, I would start adding random bits to the metadata so the hashes always change.

The only way this system even works is if there are centralized points the governments/corporations can control. Chalk this up as another point for supporting decentralization & lightweight self-hosting since it would be impossible to have oversight over such a system if anyone can spin up a personal server in their bedroom.

Yes, I agree it is dangerous. I just wanted to assess the actual threat (current and future) before jumping onto the wagon.

They say they the images are merely matched to pre-determined images found on the web. You're talking about a different scenario where AI detects inappropriate contents in an image.