One general limitation is that the decoder can only recognize concepts (“semantic features”) it has been trained for with the particular subject. The paper mentions that “we recorded responses while the subject listened to sixteen hours of naturally spoken narrative stories”. If the training material didn’t contain (say) the concept “grandmother”, the decoder probably won’t be able to decode that concept. And the more abstract the concept, the more training is probably needed. You likely also have to retrain each year or so because your brain changes.

Nevertheless, if it could somehow be commoditized into an affordable appliance, it could be very useful.