Here's a summary of how their proposed system was supposed to work: https://educatedguesswork.org/posts/apple-csam-intro/ (if you want to skim it, search for mentions of "manual review" and "visual derivative")

I suspect the key would be that there'd be a team verifying that something is actually child pornography, because the system is a perceptual hash rather than a strict comparing-bytes so before someone looks at it they're not certain.