Feel free to correct me if I'm wrong, but this is a method for decrypting _if it's matching an already known or flagged item_. It's not enabling decrypting arbitrary payloads.
From your link:
>In particular, the server learns the associated payload data for matching images, but learns nothing for non-matching images.
Past this point I'll defer to actual cryptographers (who I'm sure will dissect and write about it), but to me this feels like a decently smart way to go about this.
[1] https://www.apple.com/child-safety/pdf/CSAM_Detection_Techni...