Complexity yes, but I feel scalability is similarly a word with a misleadingly narrow connotation in practice.

Is an approach dependent on swathes of training data truly scalable if it doesn’t work for the first n attempts?