The main problem is not even the presence of machine learning content detection. It’s a lot of things predating any sort of machine learning development:

• “Screw consent” mentality: feature is by default on.
• Product tying: Apple doesn’t need to show to the user that the feature is desirable, since it’s tied to its OS.
• Leaking data: feature could be reasonably developed to run locally, but instead it sends data to someone else’s computer aka “the cloud”.