The claim of catching 99% of activity is really hard to justify. In many detection systems you cannot know your true number of false negatives. It's inherently unknowable in this case. A better metric would be mean time to detection for a given CEP sharer or perhaps a metric around how much CEP is shared before detection, or perhaps a metric for network size of CEP and how well FB can clean up an entire network at on…
Well, one place to start is with the realization that 99% of molestation comes from people the child knows. Encryption doesn't affect that pipeline at all, so the gov't shouldn't be worried.
I'd also remind that probably effectively 100% of physical child abuse and violence also comes from people the child knows.
Surveilling strangers can't have anything to do with dealing with any form of child abuse, except the extremely rare case where strangers have been able to find each other and create a subculture with which to provide evidence that exists on Facebook's servers. That's a pretty specific and narrow set of conditions.
I'm not saying Facebook allowing government surveillance on their wires can't produce some hits, but I'd be extremely curious to learn of wider implications and costs that are proportional.