In principle I can see the merit.  But in practice I see problems.  Sending huge numbers of IDs to exclude is itself expensive.  Bloom filters aren't precise.  Negentropy (the other solution often suggested) requires potentially a lot of round trips.  There are possibly other solutions like getting all the IDs first and then asking for only some of those events.  Also 256-bit IDs are cryptographically unique but we don't need cryptographic uniqueness for this avoid-double-event problem, just practical uniqueness, so using the first 64 bits of an ID to make it shorter is a reasonable optimization to whichever solution arises as the best.

I haven't heard anybody make an argument that compares the options and concludes as to which option is superior, so we are still in an exploratory stage as to this issue.  But I don't deny the problem.