Dataflow incremental refresh duplicates data

 If that's the case, how does Power BI know which records need to be updated?

It's called incremental refresh, not differential refresh. 

Incremental refresh CANNOT update records. The best you can do is flush and fill partitions.

You could use the "Detect data changes"  option but what this will do is create a shadow copy ("canary") of your semantic model, effectively doubling the storage needs.

In essence - it is YOU who needs to know which rows have changed, and who has to decide if and when to refresh the impacted partitions.