Well, Apple incorrectly thought that people would rather Apple NOT go through ALL of their photos and I think it’s an easy mistake to make.
Apple: Hey, we need to do this thing and we can either
Pore over every single photo you’ve ever taken which means we have to store your photos in a way that we can decrypt them… of COURSE meaning that if the government ever asked us to decrypt them, we can’t say we don’t have they key, so that’s a thing.
OR
We can have your device flag any potential CP (like if you have a good number of matches) ON YOUR DEVICE. That way, if you’re using iCloud, we don’t look at ANYTHING of yours unless you have CP AND we can tell anyone asking us to provide access to your photos that they’ll have to ask for your password because we don’t have any way of accessing any non-flagged images.
Which would you prefer?
Public: ENSURE THAT YOU CAN PROVIDE TO THE GOVERNMENT ANY IMAGES THAT THE GOVERNMENT REQUESTS OF YOU.
Apple: Huh… wouldn’t have figured that but… well, back to the drawing board.
I'd definitely not prefer this to be done on my device. If you upload sensitive information to a server outside your control, it's understood you don't have full control over the sensitive information anymore and therefore taking risks with the sensitive information. This risk unfortunately does not go away whether the surveillance scanning is done on device or on the server. And in actuality, when you are doing this on-device, you are unnecessarily opening up additional vectors for attack.The idea was that the hash matching is done on your phone instead of communicating with the server until some threshold of certainty was crossed. It’s a similar logic behind so much of Apple’s AI approach (do it on the native device to the amount possible instead of remote servers, in part for privacy reasons*, in part to exploit Apple’s expertise in silicon vs their relative weakness in cloud hardware).
* An example of privacy implications of cloud based CSAM scanning vs on-device: image isn’t CSAM but may be sensitive, gets uploaded to the cloud, where who knows what happens to it in addition to being hashed and compared to the CSAM database. In the Apple approach, the image doesn’t get tested by server side CSAM detection until it crosses a certain threshold of matching the hash, so the non-CSAM but sensitive information doesn’t get uploaded to the cloud unless it matches the hash to a certain threshold. There are some other theoretical aspects of on-device vs cloud based analysis. Theoretically, on device analysis is more transparent than cloud based analysis, since researchers can analyze the device doing the analysis far easier than a cloud based server. Also theoretically, it could be harder to change the hash code to match non-CSAM sensitive content unnoticed. Not CSAM matching is of course more protective of privacy than CSAM matching, but theoretically on-device CSAM matching could be more protective of privacy than cloud based CSAM matching.
It seems the biggest culprit in all this is Lindsey Graham who tried to introduce and pass the EARN IT act, trying to make providers 'earn' Section 230 immunity, and affect CALEA without having to amend it (by trying to remove E2E encryption). The wind of this most likely forced tech companies to try to pivot, causing Apple to make an unfortunate and unnecessary misstep.