Last time I checked, the preferred method was to simlify known CSAM material enough that you can hash the result, then repeat in the client end and hope nothing else has the same hash.
On the contrary, it is an argument in favour of extending ChatControl to using machine learning for detecting such cases.
The problem, again, is that we don't want to have an opaque system that can be extended to surveilling new things, because it's very difficult to audit and make sure it is not abused.