This is a misunderstanding. The goal here is not to identify new child pornography, based on ML trained models. This is to identify known child pornography, according to a hashed value. The hashed value is generated by an ML model.

So is the proposal a system for identifying known images of stolen bikes? That doesn't make sense because it isn't against the law to have an image of a stolen bike.

No. This system would only work for material that is illegal to possess. If we want to go paranoid, we could see a similar system deployed to prevent computers from having unauthorized copyrighted materials (DRM-free ebooks, academic papers, movies, etc), that’d be a horrible reality.

You recognize bikes, and compare it to a database of known stolen bikes.

CSAM only works by checking a hash. Photos of a stolen bike, especially ones that are sold online would probably have unique images taken by the thief.