What happens is that humans have intuitive coordination systems and tend to fix the same problems.
In other words, if the card title is "rename auotmerge" then both of us are likely to change it the same way, or at least either resolution is fine. In the event of a conflict we flag the multi-value in the UI so that a human user can decide if they care.
Where this approach shines is in domains where it helps avoid conflicts rather than resolving them. Naively adding replies to a comment thread might result in collisions or misordered comments across nodes of a distributed team. Automerge captures enough metadata from the operations created by users that changes "tend" to merge well.
I realize this is an argument not from first principles but from experience. We weren't able to find examples of other people building web applications with quite this approach (though we would love to hear from them if they are out there) and so the question was "will this feel alright?"
Apparently it does.
If you find this result provocative, I would encourage you to challenge it by attempting to build something with automerge and letting us know how it goes. I believe it will work well in applications where all user operations can always commute to produce a subjectively reasonable result. To make that concrete, a bank account is a poor domain to model with a tool like this, but a shared drawing is much better.