Isn’t that what they’re for?
My guess is this is because the machine learning isn’t done on the fly - it’s done in DxO’s laboratory. My understanding (feel free to correct me) is that the heuristics are built into the program and updated regularly. I don’t have an NPU, so am glad I can benefit from “AI” features by DxO, Topaz, and other developers who do it this way.
NPUs are designed to execute the calculations required to utilise existing models. The computing power required to create the models is enormous and well beyond the bounds of any single, end-user device.
NPUs are absolutely used to execute DeepPRIME variants. It seems reasonable that the AI masking calculations also use them.
I’ve read that DeepPrime used the GPU. However I’ve found the preference setting to use the NPU. So, I’m a happy camper.