I think it is a combination of both. I don't program for a living anymore, but program bugs, incomplete or incorrect data, transcription errors, malfunctioning equipment, and a whole bunch more coupled with a real lack of knowledge by both the mets and the programmers understanding of all of the realities of nature, all combine to make errors.

The models you mentioned often don't 'understand' the effects of future changes in the strength speed of change and location of data asssociated with distant but significant ridges and troughs often hundreds or even thousands of miles distant. Some do better with the future, some do great for near-term events, some do fine with 'normal' weather but as the storm gets stronger, perform poorly.

For these reasons and more, we have to look at the whole bunch of models and apply our own experience, training and common sense using the models as a tool like a mechanic uses a wrench or screwdriver....the models by themselves are useless in the same way that a wrench can't fix a car, but as a tool, both are valuable in the hands of a skilled person.