It's
better for a customer to scan images locally, on the phone, prior to upload. By doing this, the device can then encrypt images and store them in the cloud service. In this way the cloud service can be "CSAM sharing free" but never needs the symmetric keys to decrypt private images, period, for any reason.
The only thing this adds to the threat model is mistrust for false positives in scanning engine - but in even the worst case scenario here (forged false positives), you're still ahead of the Google model, where the same forged false positives would be extremely likely to result in a full account review rather than review of specific images. Everything about "the device looking at your photos" is tinfoil hat, because the device is already looking at your photos, they're decrypted in RAM! All of the threat scenarios about "Apple adds a secret government backdoor that downloads your photos and sends them to the FBI" are already equally possible today!