Even things in the neighborhood of Apple's hash and device scans can only detect existing minimally altered images and also fall down to context if they try to detect more general issues like "nude children". We don't even have to get hypothetical about false positives of anti-CSAM measures [0] they're already out there happening. The devices are blind to context and the occurrence of actual CSAM are rare (blessedly) but there's a whole genre of "my kid did something cute in the tub" that look a lot like it without the context to a machine.
[0] https://9to5google.com/2022/08/22/google-locked-account-medi...