Assignments are manual, but at least they’re not fixed. The model will still move terms into other topics (or include them in both!) if your manual suggestions are not supported by the data.
There’s usually a high level of human interpretation with topic modelling anyway. It’s very difficult to assign labels to the topics that are generated. They’re often impenetrable.
Topic modelling tends to be applied within a specific domain and analysts will typically have a lot of prior knowledge that they can inject into the process. If that leads to topics that are easier to interpret, that feels like a win to me.